Skip to content

CHANGELOG.md

Latest commit

 

History

History
287 lines (176 loc) · 22 KB

CHANGELOG.md

File metadata and controls

287 lines (176 loc) · 22 KB
Raw

Changelog

0.17.0 (2026-07-15)

⚠ BREAKING CHANGES

  • remove transcript adapter binaries (#54)

Features

  • Remove transcript adapter binaries (#54) (0f41895)
  • tui: Palette-first redesign + rebuild flows (#58) (1e0fdd6)

Bug Fixes

0.16.0 (2026-07-14)

Features

0.15.0 (2026-07-13)

⚠ BREAKING CHANGES

  • index transcripts directly, offline dedupe, background auto-index

Features

  • codex: Marketplace install, drop install.sh (#43) (179eb2c)
  • Index transcripts directly, offline dedupe, background auto-index (faeef99)

0.14.0 (2026-05-22)

Features

  • Bg codex stop hook + unified worktree paths (#36) (a713a65)

0.13.0 (2026-05-05)

Features

  • Concurrent recall + TUI activity pane (#34) (ea330f5)

0.12.0 (2026-05-01)

Features

  • Add lethe seed for history backfill (dc9b263)

0.11.0 (2026-05-01)

Features

  • Allow multiple expand at once (a47bcfb)

0.10.0 (2026-04-29)

Features

  • codex: Codex CLI plugin + plugin manifest fix (#28) (0354c5b)

0.9.0 (2026-04-28)

Features

  • retrieve: Batched union rerank, ~4x faster search --all (096c64f)

0.8.0 (2026-04-28)

⚠ BREAKING CHANGES

  • rust port — single lethe binary, bindings, polyglot release (#19)

Features

  • Add --all global search to memory-recall skill (3ccdafe)
  • Add analysis script with plots and summary table (b1e140b)
  • Add BM25 hybrid retrieval, +73% over static on LongMemEval (7dd3320)
  • Add data preparation script for NFCorpus (a47ab4b)
  • Add health metrics (diversity, anchor drift, NDCG, recall) (e0d13eb)
  • Add MemoryEntry dataclass and Tier enum (c021bfb)
  • Add Static and RandomMutation baseline stores (f20803f)
  • Clustered RIF, +5.8% NDCG — cue-dependent suppression 5x better than global (cbfe8c3)
  • Enrichment layer validated on covered queries — +8.3pp NDCG over RIF (eb4e551)
  • Exploration + rescue list with smart top-K injection (a828402)
  • Implement experiment runner with circuit breakers (4b6c4dd)
  • Implement GCMemoryStore with FAISS retrieval and GC loop (9fc0544)
  • Implement mutation and selection with unit tests (a731564)
  • Lethe tui, split recall skill, lazy session headers (#7) (b5799b3)
  • Lifecycle benchmark confirms GC tiers don't help retrieval (2bc0df3)
  • LLM enrichment layer for checkpoint 13 (pipeline, not yet run) (d9676ca)
  • MLP with per-step delta cap shows first positive result (fbd70fb)
  • Production MemoryStore API with SQLite + FAISS + BM25 (82f7076)
  • Rank-gap competition formula, clustered+gap hits +6.5% NDCG / +9.5% recall (784bd59)
  • Regex BM25 tokenizer (+3.68pp NDCG, +6.79pp Recall) (#17) (c17828c)
  • Rescue index gives +12.9% hot NDCG on LongMemEval (231e06c)
  • Retrieval-induced forgetting, first positive learned mechanism (+2.0% NDCG) (3bb484f)
  • Rust port — single lethe binary, bindings, polyglot release (#19) (8f0d1e6)
  • Scaffold repo with pyproject.toml, config, and README (f278435)
  • Ship lethe as a Claude Code plugin (87f1a40)
  • Sparse Distributed Memory (SDM) research prototype (7056b6c)
  • Stabilize MLP with lower promotion thresholds (2798a4a)
  • Switch benchmark from NFCorpus/BEIR to LongMemEval (b4a7194)
  • Tui live-expand on highlight, fix duckdb conflict and clipped search input (2159499)
  • V3 cross-encoder guided adapter mutation (019bb1f)
  • V4 learned MLP adapter + segmentation mutation (ef9a9c6)
  • V4 results on both datasets + MLP selection fix (aa4f333)
  • Wire clustered RIF into production, add video demo (d2b728b)

Bug Fixes

  • Add README to PyPI, absolute logo URL, MIT license (78188dd)
  • Average NDCG@10 over all eval queries, not single sample (413943e)
  • Enrichment rate-limiting — default concurrency 20→5, max_retries 4→8 (2c04558)
  • Lower adaptive deep-pass k_deep from 200 to 100 (#12) (5e8c1db)
  • Perf micro-wins in the index and save hot paths (#15) (a7fcf49)
  • release: Bump inter-crate version pins to 0.7.0 (c110aa1)
  • release: Drop dangling x64 refs in homebrew formula (arm64-only) (33eaa90)
  • release: Drop prepublishOnly hook so manual npm publish works (7f43025)
  • release: Install libduckdb.dylib next to lethe binaries (44f70f7)
  • release: Make cargo publish idempotent (0925ce1)
  • release: Make npm publish idempotent + ship meta package (163c61c)
  • release: Match component-prefixed tag in release.yml trigger (35ec1b7)
  • release: Re-run artifact check on bot push, prune stale wheels (b73db0e)
  • release: Scope npm under @lethe-memory, unbreak cargo publish (ee21d70)
  • release: Set User-Agent on crates.io API check (was 403'd) (c874e8e)
  • release: Switch root to release-type simple with toml updater (a3e607c)
  • Remove chunk_map.json, expand reads from DuckDB directly (5cb8c84)
  • Tui scope transitions release duckdb handles; document tui install (#10) (609acc9)

0.7.1 (2026-04-28)

Bug Fixes

  • release: Bump inter-crate version pins to 0.7.0 (c110aa1)
  • release: Drop dangling x64 refs in homebrew formula (arm64-only) (33eaa90)
  • release: Drop prepublishOnly hook so manual npm publish works (7f43025)
  • release: Install libduckdb.dylib next to lethe binaries (44f70f7)
  • release: Make cargo publish idempotent (0925ce1)
  • release: Make npm publish idempotent + ship meta package (163c61c)
  • release: Match component-prefixed tag in release.yml trigger (35ec1b7)
  • release: Scope npm under @lethe-memory, unbreak cargo publish (ee21d70)
  • release: Set User-Agent on crates.io API check (was 403'd) (c874e8e)

0.7.0 (2026-04-28)

⚠ BREAKING CHANGES

  • rust port — single lethe binary, bindings, polyglot release (#19)

Features

  • Rust port — single lethe binary, bindings, polyglot release (#19) (8f0d1e6)

Bug Fixes

  • release: Switch root to release-type simple with toml updater (a3e607c)

0.6.0 (2026-04-24)

Features

  • Regex BM25 tokenizer (+3.68pp NDCG, +6.79pp Recall) (#17) (c17828c)

0.5.1 (2026-04-24)

Bug Fixes

  • Perf micro-wins in the index and save hot paths (#15) (a7fcf49)

0.5.0 (2026-04-24)

Features

  • Tui live-expand on highlight, fix duckdb conflict and clipped search input (2159499)

0.4.2 (2026-04-24)

Bug Fixes

  • Lower adaptive deep-pass k_deep from 200 to 100 (#12) (5e8c1db)

0.4.1 (2026-04-23)

Bug Fixes

  • Tui scope transitions release duckdb handles; document tui install (#10) (609acc9)

0.4.0 (2026-04-23)

Features

  • Lethe tui, split recall skill, lazy session headers (#7) (b5799b3)

0.3.0 (2026-04-17)

Features

  • Add --all global search to memory-recall skill (3ccdafe)

0.2.2 (2026-04-17)

Bug Fixes

  • Remove chunk_map.json, expand reads from DuckDB directly (5cb8c84)

0.2.1 (2026-04-17)

Bug Fixes

  • Add README to PyPI, absolute logo URL, MIT license (78188dd)

0.2.0 (2026-04-17)

Features

  • Add analysis script with plots and summary table (b1e140b)
  • Add BM25 hybrid retrieval, +73% over static on LongMemEval (7dd3320)
  • Add data preparation script for NFCorpus (a47ab4b)
  • Add health metrics (diversity, anchor drift, NDCG, recall) (e0d13eb)
  • Add MemoryEntry dataclass and Tier enum (c021bfb)
  • Add Static and RandomMutation baseline stores (f20803f)
  • Clustered RIF, +5.8% NDCG — cue-dependent suppression 5x better than global (cbfe8c3)
  • Enrichment layer validated on covered queries — +8.3pp NDCG over RIF (eb4e551)
  • Exploration + rescue list with smart top-K injection (a828402)
  • Implement experiment runner with circuit breakers (4b6c4dd)
  • Implement GCMemoryStore with FAISS retrieval and GC loop (9fc0544)
  • Implement mutation and selection with unit tests (a731564)
  • Lifecycle benchmark confirms GC tiers don't help retrieval (2bc0df3)
  • LLM enrichment layer for checkpoint 13 (pipeline, not yet run) (d9676ca)
  • MLP with per-step delta cap shows first positive result (fbd70fb)
  • Production MemoryStore API with SQLite + FAISS + BM25 (82f7076)
  • Rank-gap competition formula, clustered+gap hits +6.5% NDCG / +9.5% recall (784bd59)
  • Rescue index gives +12.9% hot NDCG on LongMemEval (231e06c)
  • Retrieval-induced forgetting, first positive learned mechanism (+2.0% NDCG) (3bb484f)
  • Scaffold repo with pyproject.toml, config, and README (f278435)
  • Ship lethe as a Claude Code plugin (87f1a40)
  • Sparse Distributed Memory (SDM) research prototype (7056b6c)
  • Stabilize MLP with lower promotion thresholds (2798a4a)
  • Switch benchmark from NFCorpus/BEIR to LongMemEval (b4a7194)
  • V3 cross-encoder guided adapter mutation (019bb1f)
  • V4 learned MLP adapter + segmentation mutation (ef9a9c6)
  • V4 results on both datasets + MLP selection fix (aa4f333)
  • Wire clustered RIF into production, add video demo (d2b728b)

Bug Fixes

  • Average NDCG@10 over all eval queries, not single sample (413943e)
  • Enrichment rate-limiting — default concurrency 20→5, max_retries 4→8 (2c04558)

Documentation

  • Checkpoint 14 negative result — retrieval-only ceiling reached (ca6fb06)
  • Checkpoint 15 — SDM prototype, negative result (d657de4)
  • Checkpoints 16-17 + benchmark-methodology disclaimer (082b0c6)
  • Restructure — move result snapshots into benchmarks/results, rewrite BENCHMARKS and RESEARCH_JOURNEY (1373264)
  • Rewrite README for checkpoints 13+17 and methodology note (7637ce2)