notes.md

Infinity Context LoCoMo Retrieval Quality Notes

2026-07-02

• Baseline focused memory-comparison checks passed before changes: uv run --extra dev pytest -q tests/unit/test_memory_comparison_bundle_planner.py tests/unit/test_memory_comparison_rerank_policy.py tests/unit/test_memory_comparison_answer_context.py tests/unit/test_memory_comparison_quality_support_gaps.py tests/unit/test_memory_comparison_query_role_diagnostics.py -> 163 passed.
• Implemented typed favorite_preference support so favorite/favourite questions require explicit favorite evidence instead of being satisfied by generic preference/like evidence.
• Added favorite_support across query planning, evidence bundle planning, typed rerank support, answer-context backfill matching, and diagnostics.
• Tightened the missing favorite evidence safety cap so generic preference matches cannot outrank explicit favorite evidence solely because of upstream retrieval score.
• Updated stale query-role expectation for future home-move/current-goal decomposition to reflect typed current_goal_support.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 495 passed, 1 warning.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json is absent. Fast-readiness blockers were otherwise empty; no long/full LoCoMo run was attempted.
• git push origin main -> blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 53

• Added typed date-profile support for birthdate and birth-date questions such as "What is Alex's birthdate?"
• Added explicit birthdate evidence surfaces while keeping birth-certificate wording guarded as a topical distractor.
• Preserved the existing temporal/date bundle behavior for date-looking birthdate questions while requiring explicit date-profile evidence.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_date_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_birthdate_date_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_dob_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_date_profile_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 549 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 58

• Expanded typed diet-profile intent for "What can X not eat?" questions where the person appears between can and not eat.
• Added matching diet evidence support for explicit can not eat statements.
• Added guarded rerank coverage so direct cannot-eat evidence outranks topical restaurant mentions.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_diet_avoidance_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_cannot_eat_diet_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_diet_avoidance_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_diet_profile_ignores_diet_book_wording -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 558 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 59

• Expanded typed preference support for prefer questions such as "What tea does Alex prefer?"
• Added tea/coffee preference query and evidence terms while preserving the existing favorite/go-to preference behavior.
• Added guarded rerank coverage so direct preferred-tea evidence outranks a topical tea purchase mention.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_favorite_preference_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_prefer_tea_preference_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_favorite_preference_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_go_to_favorite_preference_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 559 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 57

• Guarded activity-profile questions with "free time" or "pastime" so they no longer request unrelated temporal support just because the word time appears.
• Expanded free-time activity query terms through the existing hobby/activity path so direct evidence like "In my free time, I enjoy painting" receives typed activity_support.
• Added guarded rerank coverage so direct free-time activity evidence outranks topical free-time-management mentions.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_activity_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_free_time_activity_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_activity_profile_evidence -> 3 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 557 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 56

• Expanded typed location support for LoCoMo-style current residence phrasing: "Where is Alex living?", "Which city is Alex based in?", "What is Alex's current city?", and "Where is Alex staying?"
• Added grounded current-city/home/location evidence surfaces and location boost support for stemmed living evidence.
• Kept contact address questions on the contact-profile path and stay-present wording off the location path.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_location_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_current_city_location_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_staying_location_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_location_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_based_location_profile_evidence -> 5 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policies.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 556 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 55

• Expanded typed employment-profile intent for employer questions such as "Who is Alex's employer?" and "What employer does Alex work for?"
• Kept conversational action questions like "What did Alex discuss with his employer?" on the communication/action path instead of requiring employment profile evidence.
• Added guarded rerank coverage so direct employer evidence outranks topical company mentions.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_employment_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_employer_employment_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_employment_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_occupation_employment_profile_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 554 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 54

• Expanded typed status-profile intent for family relationship questions such as "Does Alex have children?", "Who are Alex's parents?", and "What is Alex's wife's name?"
• Added explicit family status evidence support for surfaces like "I have two children" and "my parents" while preserving project-partner guards.
• Added guarded rerank coverage so direct children/wife evidence outranks topical family-word mentions.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_family_status_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_children_status_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_wife_name_status_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_status_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_possessive_status_evidence -> 5 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 553 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 24

• Added typed date-profile planning for explicit wedding-date questions by mapping "wedding date" to the existing anniversary/date support path.
• Kept wedding venue/action wording out of date-profile classification.
• Tightened the missing date-profile evidence safety cap so topical wedding or date mentions do not outrank direct wedding-date evidence solely from initial retrieval score.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_date_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_wedding_date_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_married_anniversary_date_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_birthdate_date_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_birthday_day_date_profile_evidence -> 6 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policy.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 550 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 52

• Expanded typed status-profile intent for married-to and marital-status questions such as "Who is Alex married to?"
• Added a project-partner guard so "Who did Alex partner with on the project?" does not receive typed relationship/status evidence requirements.
• Added guarded rerank coverage so explicit married-to status evidence outranks topical wedding mentions.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_handles_question_bound_partner_status_terms tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_married_to_status_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_status_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_possessive_status_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 548 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 51

• Added typed alias-profile support for middle/full/legal name questions such as "What is Alex's middle name?"
• Preserved name-attribute terms in compact alias query planning only for explicit name-attribute questions, keeping ordinary nickname/go-by queries stable.
• Added guarded rerank coverage so explicit middle-name evidence outranks topical uses of "name" and "middle."

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_alias_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_middle_name_alias_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_alias_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_go_by_alias_profile_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 547 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 50

• Added typed education-profile support for graduation questions such as "Where did Alex graduate from?"
• Prevented graduation-from wording from being misrouted as origin/location evidence while keeping ordinary origin questions unchanged.
• Added guarded rerank coverage so explicit graduation-school evidence outranks topical "from Stanford" mentions, and non-academic "graduate to doing" wording does not receive typed education boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_education_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_graduation_education_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_degree_education_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_major_education_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_education_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_school_education_evidence -> 6 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 546 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 49

• Tightened degree-question education intent so topical wording such as "degree of difficulty" does not create an education evidence need.
• Added focused degree education rerank coverage for explicit degree evidence such as "I earned a degree in psychology."
• Kept academic degree questions using compact education query terms while guarding non-academic degree mentions from typed education boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_education_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_degree_education_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_major_education_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_education_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_school_education_evidence -> 5 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 545 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 48

• Tightened education-profile intent for major/degree questions so topical wording such as "major issue" does not create an education evidence need.
• Preserved major/degree terms in compact education query planning only for academic-major or degree questions.
• Added education-profile evidence surfaces for explicit major and degree statements while keeping topical major/degree mentions from receiving typed education boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_education_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_major_education_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_education_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_school_education_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 544 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 47

• Expanded typed employment-profile support for salary, wage, pay-rate, and hourly-rate questions such as "What is Alex's salary?"
• Preserved compensation terms in compact employment query planning only for compensation questions, keeping ordinary job/company/workplace queries stable.
• Added guards so salary-cap or other topical compensation wording does not receive typed employment evidence boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_employment_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_salary_employment_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_employment_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_occupation_employment_profile_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 543 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 46

• Expanded typed skill-profile support for certification and credential questions such as "What certification does Alex have?"
• Added certification/credential surfaces to focused skill query planning without changing language or instrument query behavior.
• Added guards so birth-certificate wording does not receive typed skill evidence boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_skill_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_certification_skill_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_skill_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_fluent_language_skill_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 542 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 45

• Expanded typed pet-profile support for pet microchip-number questions such as "What is Alex's dog's microchip number?"
• Added microchip/number to focused pet query planning only when the question asks for a pet microchip, keeping ordinary pet-name and breed queries stable.
• Added guards so topical chip wording does not receive typed pet evidence boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_pet_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_pet_microchip_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_pet_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_pet_breed_profile_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 541 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 44

• Expanded typed vehicle-profile support for license-plate questions such as "What is Alex's license plate?"
• Preserved license/licence/plate surfaces in compact vehicle-support query planning only for license-plate questions, keeping ordinary car/model/color vehicle queries unchanged.
• Added guards so topical dinner-plate wording does not receive typed vehicle evidence boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_vehicle_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_license_plate_vehicle_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_vehicle_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_vehicle_model_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 540 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 43

• Expanded typed health-profile support for blood-type questions such as "What is Alex's blood type?"
• Preserved blood/type surfaces in compact health-support query planning only for blood-type queries, keeping ordinary doctor/medication/allergy health queries unchanged.
• Added guards so topical blood-drive/type wording does not receive typed health evidence boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_health_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_blood_type_health_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_health_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_primary_care_physician_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 539 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 42

• Expanded typed contact-profile support for emergency-contact questions such as "Who is Alex's emergency contact?"
• Added focused compact contact query planning that preserves the emergency surface without degrading ordinary phone/number and social-contact queries.
• Added guards so event questions like "Who did Alex contact during the emergency?" and topical emergency-contact distractors do not receive typed contact evidence boosts.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_emergency_contact_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_emergency_contact_evidence tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_contact_number_queries tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_social_contact_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_social_contact_evidence -> 5 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 538 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 41

• Expanded typed contact-profile support for social contact handles and platform numbers, including Instagram/Telegram/Signal/Slack/WhatsApp handle, username, and number questions.
• Prioritized platform/handle terms in compact contact-support query planning so "How can I reach Alex on Instagram?" retrieves focused social contact evidence instead of only broad phone/email terms.
• Added guards so topical platform mentions and unrelated "handle" wording do not receive typed contact evidence.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_social_contact_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_social_contact_evidence tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_contact_number_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_contact_number_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_reach_contact_evidence -> 5 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 536 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 40

• Routed "How can I contact/reach X?" questions through typed contact-profile support instead of generic how/causal retrieval.
• Added explicit reach/contact-at evidence support for phone and email surfaces, such as "You can reach me at 555-0101."
• Kept non-contact reach questions, such as reaching a summit, on the existing how/causal path and added regressions for both the positive and guard cases.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_contact_number_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_contact_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_contact_number_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_reach_contact_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_contact_profile_ignores_addressing_issue_wording -> 5 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 534 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 39

• Routed "date of birth" and "DOB" questions through typed date-profile support instead of generic temporal/single-fact retrieval.
• Added explicit DOB evidence detection and date-profile surface grounding so evidence like "My DOB is May 5" receives typed date_support.
• Prevented uppercase DOB from being treated as a person/entity during query planning, and added regressions proving DOB evidence outranks topical birth certificate distractors.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_date_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_dob_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_born_date_profile_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_candidate_features.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 533 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 38

• Expanded typed date-profile evidence for anniversary answers phrased as marriage-date facts, such as "We got married on July 2."
• Added wedding/married/marry grounding terms to the date-profile category for anniversary questions so typed date_support can distinguish explicit anniversary evidence from topical wedding mentions.
• Allowed direct, localized typed relation evidence with full category coverage to use a higher boost cap, preventing upstream-scored distractors that still miss required profile evidence from edging out grounded typed evidence.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_date_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_married_anniversary_date_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_birthday_day_date_profile_evidence -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policy.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 532 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 37

• Expanded typed date-profile intent for birthday/anniversary questions phrased as "what/which day" or "what month", such as "What day is Alex's birthday?"
• Kept existing birthday duration and birthday-gift guards intact, so duration questions and memento questions do not get promoted to date-profile evidence.
• Added query decomposition and rerank regressions proving day/month date questions request date_support and explicit birthday-date evidence outranks topical birthday-gift distractors.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_temporal_action_queries tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_date_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_birthday_day_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_born_date_profile_evidence -> 5 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 531 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 36

• Expanded typed alias-profile support for "go by" questions such as "What name does Alex go by?" and evidence such as "I go by Sunny."
• Added "go by" to alias query fanout and guarded typed grounding so proper alias surfaces can be boosted without treating route/location "went by" mentions as alias evidence.
• Added query decomposition and rerank regressions proving go-by alias evidence receives typed alias_support while a pharmacy route distractor remains untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_alias_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_alias_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_go_by_alias_profile_evidence -> 3 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_candidate_features.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 530 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 35

• Expanded typed age-profile support for age evidence phrased as turning or having turned an age, such as "I'm turning 32 next month."
• Added age query fanout for the normalized turn surface while retaining birthday/born/old terms for existing age questions.
• Added a rerank regression proving turning-age evidence receives typed age_support while a 32-page book distractor remains untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_age_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_age_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_turning_age_profile_evidence -> 3 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 529 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 34

• Expanded typed favorite-preference support for "go-to" favorite questions such as "What is Alex's go-to restaurant?" and evidence such as "My go-to restaurant is Cafe Luna."
• Routed hyphenated go-to questions through the existing favorite-support query role while keeping ordinary "go to" education/travel wording on its existing paths.
• Added query decomposition and rerank regressions proving go-to evidence receives typed favorite_support while topical restaurant mentions remain untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_favorite_preference_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_favorite_preference_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_go_to_favorite_preference_evidence -> 3 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 528 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 33

• Expanded typed diet-profile support for avoidance evidence such as "I avoid seafood" and common restrictions including seafood, eggs, soy, and lactose.
• Added guarded diet query routing for "What food does X avoid?" while keeping diet-book wording out of typed diet support.
• Added query decomposition and rerank regressions proving dietary avoidance evidence receives typed diet_support while topical seafood mentions remain untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_diet_avoidance_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_diet_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_diet_avoidance_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_diet_profile_ignores_diet_book_wording -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 527 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 32

• Expanded typed location support for "Where is X based?" questions and evidence such as "I am based in Denver now."
• Added based-location query terms and location evidence scoring while guarding non-location "based on" questions from location-support routing.
• Added query decomposition and rerank regressions proving based-location evidence receives location_support and topical location mentions remain untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_location_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_location_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_based_location_profile_evidence -> 3 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policies.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 525 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 31

• Expanded typed contact-profile support for contact-number questions such as "What is Alex's number?" and evidence such as "My number is 555-0101."
• Kept number handling guarded so non-contact questions like "What number did Alex pick?" do not receive contact-profile query roles.
• Added query decomposition and rerank regressions proving contact-number evidence receives typed contact_support while topical number mentions stay untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_contact_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_contact_number_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_contact_number_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_contact_profile_ignores_addressing_issue_wording -> 4 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 524 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 30

• Expanded typed status-profile support for extended kinship questions such as "Who is Dana's cousin?" without widening generic relationship-status fanout.
• Added cousin/grandparent relation terms and focused cousin query variants so family-role questions can retrieve typed status evidence while topical cousin mentions remain untyped.
• Added query decomposition and rerank regressions proving cousin evidence receives typed status_support.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_handles_question_bound_person_role_terms tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_status_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_extended_kinship_status_evidence -> 3 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 522 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 29

• Expanded typed health-profile support for primary-care physician questions and evidence such as "My primary care physician is Dr. Lee."
• Added physician to health query-role fanout and typed relation category terms, while keeping evidence support grounded on profile phrasing like primary care physician or appointments so topical physician mentions do not receive health-profile evidence.
• Added query decomposition and rerank regressions proving primary-care physician evidence receives typed health_support while a topical physician article remains untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_health_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_health_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_primary_care_physician_evidence -> 3 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 521 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 28

• Expanded typed vehicle-profile evidence for owned model shorthand such as "My Tesla is blue" and person-possessive model evidence such as "Alex's Tesla is blue," which LoCoMo-style car questions should treat as vehicle evidence even when the word "car" is absent.
• Kept vehicle model grounding out of compact query fanout: possessive model surfaces now satisfy the vehicle category grounding gate without adding brand/model terms to search queries.
• Added rerank regressions proving owned/possessive Tesla evidence receives typed vehicle_support while topical Tesla mentions remain untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_vehicle_model_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_person_possessive_vehicle_model_evidence -> 2 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_candidate_features.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 520 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 27

• Expanded typed pet-profile support for breed questions such as "What breed is Alex's dog?" and owned/named breed evidence such as "My golden retriever is named Luna."
• Kept breed terms conditional in compact pet query fanout so ordinary pet-name questions preserve their previous focused query terms.
• Added a rerank regression proving owned golden-retriever evidence receives typed pet_support while a topical golden-retriever park sighting remains untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_pet_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_pet_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_pet_breed_profile_evidence -> 3 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 518 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 26

• Expanded typed employment-profile evidence for explicit occupation identity turns such as "I'm a nurse" instead of requiring "work for/at/as" or "job is" wording.
• Kept occupation grounding out of query fanout: explicit occupation profile surfaces now satisfy the employment category grounding gate without adding a broad occupation list to compact search queries.
• Added a rerank regression proving "I'm a nurse at the clinic" receives typed employment_support and outranks a higher-scored topical nurse appointment mention.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_employment_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_employment_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_occupation_employment_profile_evidence -> 3 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 517 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_candidate_features.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 25

• Expanded location-origin support for "where was X raised" questions and "raised in PLACE" evidence.
• Guarded raised-location questions from picking up generic charity/fundraising raise relation fanout, so location support queries stay focused on origin evidence instead of awareness/fundraiser terms.
• Added regressions proving raised-in-Toronto evidence receives location_transition support while a Toronto fundraising distractor stays untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_location_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_raised_origin_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_origin_profile_evidence -> 3 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 516 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policies.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 24

• Expanded typed skill-profile language support for bilingual wording in both query planning and evidence detection.
• Prioritized language ability terms in skill support fanout so know, fluent, and bilingual questions keep the matching term in the compact query instead of being crowded out by lower-priority instrument variants.
• Added regressions proving bilingual-in-Spanish evidence receives typed skill_support while a topical Spanish cookbook mention stays untyped.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_skill_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_fluent_language_skill_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_bilingual_language_skill_evidence -> 3 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 515 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> blocked because the non-interactive runtime has no GitHub username/credential prompt available.

Next Steps

• Continue adding typed support for narrow LoCoMo facets where generic categories can over-admit evidence.
• Re-run locomo-fast only after dataset/auth/service preflight is green.

2026-07-02 Follow-up

• Promoted date_profile and status_profile relation facets into explicit compact query roles (date_support, status_support) instead of leaving their focused lexical fanout under generic/inference roles.
• Added assertions that birthday/anniversary and relationship/kinship queries carry the typed support roles in the query plan.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 495 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py tests/unit/test_memory_comparison_benchmark.py -> passed.

2026-07-02 Follow-up 2

• Added favorite_support to query-plan integrity diagnostics so fast gates can report missing typed favorite fanout instead of treating any selected query family as acceptable.
• Verified the expanded memory-comparison test set after adding typed date, status, and favorite query-plan diagnostics.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 496 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_quality_diagnostics.py tests/unit/test_memory_comparison_benchmark.py tests/unit/test_memory_comparison_quality_diagnostics.py -> passed.

2026-07-02 Follow-up 3

• Tightened answer-context backfill role matching: typed query roles such as favorite_support no longer count as missing-role support unless the candidate also has the matching evidence category/content signal.
• Added a regression test where generic preference evidence retrieved by a favorite_support query loses to explicit favorite evidence and is not marked as satisfying the missing favorite role.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 497 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_answer_context_backfill.py tests/unit/test_memory_comparison_answer_context.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 19

• Expanded status-profile relation support to recognize named possessive person-role evidence such as "Riley is Dana's roommate." This improves LoCoMo-style kinship/roommate/colleague questions where evidence is phrased with names rather than first-person pronouns.
• Added a rerank regression proving a topical roommate mention is not treated as status evidence while the named possessive status turn receives typed status support and ranks first.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_possessive_status_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_status_profile_evidence -> 2 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 511 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> blocked because the non-interactive runtime has no GitHub username/credential prompt available.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 20

• Tightened named status relation matching so possessive role phrases require a named person relation surface such as "Riley is Dana's roommate" or "Dana's roommate is Riley." A topical phrase like "Dana's roommate matching app" no longer satisfies status_profile.
• Expanded the named-status rerank regression with a higher-scored named app distractor to prove it is not treated as typed status evidence.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_possessive_status_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_status_profile_evidence -> 2 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 511 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 21

• Expanded date-profile relation support for birth-date evidence phrased as "I was born May 5" or "date of birth is May 5." This lets birthday/date questions use typed date_support even when the evidence does not repeat the word "birthday."
• Added a rerank regression proving a topical birthday-gift mention stays untyped while birth-date evidence receives typed date support and ranks first.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_born_date_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_date_profile_evidence -> 2 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 512 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 22

• Expanded education-profile relation support for named school evidence such as "I go to Stanford." School questions can now use typed education_profile support even when the evidence names the institution without repeating "school" or "university."
• Added a rerank regression proving a topical Stanford mention stays untyped while the named-school education evidence receives typed relation support and ranks first.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_named_school_education_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_education_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_education_profile_queries -> 3 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 513 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_candidate_features.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 18

• Allowed answerability boosts for grounded typed category evidence even when the category detector is the main evidence signal and relation-token hits are sparse. This helps typed profile/action evidence rank by direct answerability instead of relying only on lexical density.
• Preserved the stricter communication grounding rule: communication answerability still requires speaker grounding when the query names a speaker, so recipient-side turns do not outrank the actual speaker turn.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_rerank_policy.py -k "answerability" -> 4 passed, 36 deselected.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_speaker_grounded_communication_evidence -> 1 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 510 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policies.py tests/unit/test_memory_comparison_rerank_policy.py -> passed.

2026-07-02 Follow-up 4

• Tightened incomplete-bundle answer-context backfill further: when required roles are missing, retrieval backfill now excludes candidates that do not satisfy any missing role with matching evidence. This prevents generic/noise retrieval items from being appended just because the backfill target count has spare room.
• Updated answer-context tests so generic preference, metadata-only temporal, visual-only temporal and unrelated noise candidates are excluded from role-repair backfill.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 497 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_answer_context_backfill.py tests/unit/test_memory_comparison_answer_context.py -> passed.

2026-07-02 Follow-up 5

• Surfaced favorite_support_count as a first-class evidence bundle and answer-context diagnostic instead of only burying it under generic typed relation counts.
• Added a regression that a required favorite bundle role is satisfied only by explicit favorite_preference evidence, while generic preference evidence is rejected for that role even when it has stronger bundle score.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 498 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_bundle_planner.py packages/infinity_context_server/infinity_context_server/memory_comparison_answer_context.py tests/unit/test_memory_comparison_bundle_planner.py tests/unit/test_memory_comparison_answer_context.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 17

• Tightened answer-context backfill for the new typed action_support role. Retrieval backfill now requires matching action_event evidence before an action-support candidate can repair a missing bundle role; merely arriving from an action_support query role is not enough.
• Added a regression where a stronger query-role-only action candidate is excluded and explicit action evidence is backfilled.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_answer_context.py -k "action_role or backfill_requires" -> 4 passed, 18 deselected.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 509 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_answer_context_backfill.py tests/unit/test_memory_comparison_answer_context.py -> passed.

2026-07-02 Follow-up 6

• Propagated typed relation support totals and per-role counts from evidence bundle quality into answer-context diagnostics and aggregate metrics. This gives fast gates visibility into selected health/date/status/education/ employment/favorite/skill/vehicle/pet support instead of only generic bundle source information.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 498 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_answer_context.py tests/unit/test_memory_comparison_answer_context.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 7

• Added a bundle-quality component score for grounded typed relation support so non-favorite profile evidence such as health/date/status/education/ employment/skill/vehicle/pet support contributes directly to bundle confidence instead of only appearing in reason codes.
• Kept favorite support on its dedicated component to avoid double-counting the already first-class favorite evidence score.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 498 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_bundle_planner.py tests/unit/test_memory_comparison_bundle_planner.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 8

• Relaxed typed profile rerank grounding for category detectors: health/date/ status/education/employment/favorite-style profile hits can now receive typed relation support when provenance and entity/speaker grounding are present, even if no separate relation token hit was extracted.
• Kept non-profile categories such as causal/support-goal on the stricter relation-surface path to avoid promoting broad conversational reactions.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 500 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policies.py tests/unit/test_memory_comparison_rerank_policy.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 9

• Tightened typed relation rerank role boosts so the query-role bonus is based on typed roles with matching category hits, not every requested typed support role. This avoids over-boosting mixed profile queries when only one typed facet is actually evidenced.
• Added diagnostics for benchmark_typed_relation_support_hit_roles so fast reports can distinguish requested typed roles from grounded typed hits.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 501 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policies.py tests/unit/test_memory_comparison_rerank_policy.py -> passed.

2026-07-02 Follow-up 10

• Added query-role effectiveness diagnostics for typed relation hit roles so fast reports can distinguish a typed query role that merely appeared on a lifted candidate from a typed role with matching evidence.
• Added per-role typed relation hit counts/rates and a roles_without_typed_relation_hits list to expose mixed profile fanout where only part of the requested typed evidence was actually grounded.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 502 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_quality_query_roles.py tests/unit/test_memory_comparison_quality_diagnostics.py -> passed.

2026-07-02 Follow-up 11

• Promoted typed relation hit-role gaps into fast-gate query-role breakdowns. A lifted or selected typed query role now still reports a gap when no matching typed evidence category was actually hit.
• Added fast-gate coverage for mixed health/status profile fanout where only health evidence is grounded, so status support is reported as typed_relation_not_hit.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 503 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_quality_diagnostics.py tests/unit/test_memory_comparison_quality_diagnostics.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 12

• Made query-role gaps part of fast-gate readiness. A fast/preflight report now fails ready_for_full_locomo when selected retrieval has query-role gaps, including typed relation roles that were requested but did not produce a matching typed evidence hit.
• Added assertions that both a normal query-role selection gap and a mixed health/status typed-hit gap fail the new query_role_gaps_clear gate.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 503 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_quality_diagnostics.py tests/unit/test_memory_comparison_quality_diagnostics.py -> passed.

2026-07-02 Follow-up 13

• Tightened typed-hit gap diagnostics so roles_without_typed_relation_hits only considers roles from the typed relation support registry. Non-typed relation roles such as preference_support no longer create false typed-hit readiness failures.
• Added diagnostics and fast-gate regressions proving preference support can remain clear without typed relation hit metadata, while typed profile roles still report typed-hit gaps.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 505 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_quality_query_roles.py tests/unit/test_memory_comparison_quality_diagnostics.py -> passed.

2026-07-02 Follow-up 14

• Added lifted-only answerability gap diagnostics to the fast gate so reports separate general missing-evidence candidates from candidates the reranker actually boosted or otherwise lifted.
• Added a lifted_answerability_gaps_clear readiness gate. Full LoCoMo is now blocked when a lifted candidate still carries missing_*_evidence reason codes, while non-lifted distractor gaps remain diagnostic-only.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_quality_diagnostics.py -k "answerability or ready_for_full or lifted" -> 3 passed, 43 deselected.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 506 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_quality_diagnostics.py tests/unit/test_memory_comparison_quality_diagnostics.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 15

• Made fast-gate readiness fail when a query plan omits a query family required by the evidence roles. This uses the existing missing_evidence_role_query_family_total signal instead of gating all dropped/fanout plan diagnostics.
• Added an otherwise-ready favorite-support regression proving a base-only query plan now fails query_plan_evidence_roles_clear.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_quality_diagnostics.py -k "query_plan or ready_for_full" -> 11 passed, 36 deselected.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 507 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_quality_diagnostics.py tests/unit/test_memory_comparison_quality_diagnostics.py -> passed.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 16

• Added typed action_support for concrete "what did X take/send/paint/book/etc." action questions that are not better handled by visual evidence. This gives LoCoMo-style action asks a compact role, query fanout, typed relation hit, bundle selection path, fusion weight, and fast-gate diagnostics.
• Kept visual picture/photo/image/video questions on the existing visual path so action support does not over-boost media evidence questions.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 508 passed, 1 warning.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_policies.py packages/infinity_context_server/infinity_context_server/memory_comparison_bundle_planner.py packages/infinity_context_server/infinity_context_server/memory_comparison_quality_support.py packages/infinity_context_server/infinity_context_server/memory_comparison_quality_diagnostics.py packages/infinity_context_server/infinity_context_server/memory_comparison_quality_query_roles.py packages/infinity_context_server/infinity_context_server/memory_comparison_candidate_fusion.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py tests/unit/test_memory_comparison_bundle_planner.py -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.
• git push origin main -> still blocked because the non-interactive runtime has no GitHub username/credential prompt available.

2026-07-02 Follow-up 23

• Expanded typed skill-profile support for language ability questions phrased as "what language does X know" or "what language is X fluent in" instead of only "speak" wording.
• Added fluent/know language support to skill query fanout and rerank evidence detection while keeping instrument questions on the existing instrument/play query terms.
• Added regressions proving fluent-in-Spanish evidence receives typed skill_support and a topical Spanish-food distractor does not.

Verification

• uv run --extra dev pytest -q tests/unit/test_memory_comparison_benchmark.py::test_query_decomposition_expands_skill_profile_queries tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_skill_profile_evidence tests/unit/test_memory_comparison_benchmark.py::test_benchmark_rerank_boosts_fluent_language_skill_evidence -> 3 passed, 1 warning.
• uv run --extra dev pytest -q tests/unit/test_memory_comparison*.py -> 514 passed, 1 warning.
• uv run --extra dev ruff check packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_text.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank.py packages/infinity_context_server/infinity_context_server/memory_comparison_intent.py packages/infinity_context_server/infinity_context_server/memory_comparison_rerank_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_query_terms.py packages/infinity_context_server/infinity_context_server/memory_comparison_relation_support.py tests/unit/test_memory_comparison_benchmark.py -> passed.
• uv run --extra dev pytest -q tests/architecture/test_memory_boundaries.py -> 6 passed.
• git diff --check -> passed.
• uv run --extra dev python -m infinity_context_server.eval memory-comparison-benchmark --dataset ./datasets/locomo10.json --memo-api-url http://127.0.0.1:7788 --mem0-url http://127.0.0.1:8888 --benchmark locomo --locomo-ingest-mode official-turns --case-set locomo-fast --report-mode compact --top-k 200 --top-k-cutoff 10 --top-k-cutoff 20 --top-k-cutoff 50 --top-k-cutoff 200 --allow-live --preflight-only -> blocked safely because ./datasets/locomo10.json and memory auth token are absent. Fast-readiness blockers were empty; no long/full LoCoMo run was attempted.