You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
| Governed retrieval | Retrieval returns Context Packages, not loose chunks, writes explicit retrieval-plan metadata, applies structured subject/action/object and asset filters, filters Facts, Memory Objects, and asset extraction projections by partition/security scope before package assembly, marks affected packages stale when returned Facts or Memory Objects are superseded, can refresh a stale package from its original request scope, can batch refresh stale packages, and exposes API, CLI/cron, and supervised scheduler triggers for app/agent/runtime workflows. |
112
112
| Active Memory Pools | Task-scoped working memory can load Context Packages, refresh stale loaded Context Packages, and publish observations as pending Claims. |
113
-
| Smart memory intake |`Memory.remember/2` can gate low-salience writes, skip semantic duplicates, update superseded memories, and attach intake metadata before the governed Source Package and pending Claim bridge runs. |
113
+
| Smart memory intake |`Memory.remember/2`and `POST /api/memory/remember`can gate low-salience writes, skip semantic duplicates, update superseded memories, and attach intake metadata before the governed Source Package and pending Claim bridge runs. |
114
114
| Tool/model governance | Registered tools and model operations can enforce privileges, partitions, required inputs, required outputs, audit links, and the first governed execution path. |
115
115
| Connector governance | Connector sync runs through the governed tool-call surface by default, blocking unauthorized runs before connector execution and recording both connector-run and tool-call audit rows when allowed. Raw sync requires an explicit `governed: false` bypass. |
116
116
| Evaluation runner | Benchmark/evaluation runs can execute against governed retrieval, load JSON/JSONL datasets, assemble Context Packages per question, produce deterministic local answer surfaces, judge expected-answer matches, persist per-case scores, and update aggregate run scores. `mix optimal.eval.run` exposes the flow for CLI/cron use, and external answerer/judge callbacks can be plugged in later without changing the storage contract. |
@@ -136,7 +136,7 @@ mix test test/memory_core/spine_test.exs test/pipeline/multimodal_adapter_runner
136
136
28 tests, 0 failures
137
137
138
138
mix test test/api/router_test.exs --seed 0
139
-
27 tests, 0 failures
139
+
31 tests, 0 failures
140
140
141
141
mix test test/connectors/asset_ingest_test.exs --seed 0
0 commit comments