agent-working-memory 0.17.0 → 0.19.0

package/README.md

@@ -42,7 +42,7 @@ including the ones we think are weakest.
 
 | | |
 |---|---|
-| **Returns the right memory first** | **93.0%** of identifier queries · **92.2%** of topic queries · **97.0%** in the top five |
+| **Returns the right memory first** | **92.7%** of identifier queries · **92.7%** of topic queries · **97.3%** in the top five |
 | **Stays silent when it should** | **90.0%** correct abstention on questions about facts never stored |
 | **Costs the same at any scale** | Scoped recall answers in **~630 tokens flat**. Carrying the store instead: ~1.3M tokens |
 | **Lets a cheap model punch up** | Small model + AWM beat a frontier model on a 15-task domain workload: **14/15 vs 7/15**, at ~1/40th the cost |
@@ -181,10 +181,13 @@ compatible, and the store is the same SQLite file before and after.
 
 ---
 
-## What's new — v0.17.0
+## What's new — v0.19.0
 
 | | |
 |---|---|
+| **0.19.0** | From the speed and memory review: a warm recall was 612 ms and 474 ms of it was the cross-encoder. The **q8 reranker is the default** — rerank 35–45% faster in an interleaved in-process comparison, and on the full suites identifier 92.7% / topic 92.7% first-result (fp32: 93.0 / 92.2), top-5 and abstention unchanged. The **shared model daemon is on by default** for `awm mcp` and `awm serve`, ending the ~4 s cold first recall. The **prime hook caches** its answer for the same question for a minute instead of recalling on every prompt. SQLite gets a 64 MB page cache and mmap. The bench and profile scripts now measure in-process. |
+| **0.18.1** | Two defects MWA found within hours of the 0.18 defaults. A bare month after a preposition ("in October") now counts as a date for the conflict detector, so three versions of one fact are marked with newer/older. Two attempts to make the engine put the newest first were **measured and rejected**: a reranker-score epsilon cost identifier s@1 93.0% → 78.7%, and reordering on conflict marks cost 93.0% → 74.7% with s@5 unchanged. Marks are a cue to check the source, not a verdict on rank; `newestFirstAmongConflicts` is exported for harnesses whose policy is "newer beats older". Ranking is unchanged from 0.18.0. |
+| **0.18.0** | A superseded memory could **outrank its successor** whenever `AWM_RERANK2=1` — the setting `awm setup` writes everywhere — because the second-stage re-sort ignored the supersession penalty; fixed, with the test run reranker-on. `valid_to` now actually expires (it was display-only; the live store served 482 expired memories as current). Every recall line carries its **written date**, two versions of one fact are **marked as a possible conflict** with the newer one named, recall **confidence** is printed on every reply, and when the question asks for a time, date, amount, count or status that no returned memory holds, the reply says so instead of leaving the gap. `memory_write` takes **`correction: true`** to supersede by intent without the old id. The three measured rerank wins are **code defaults** (`=0` / `=prefix` to disable). A less-trusted write (`trust=` tag) **never merges into or retires a more-trusted memory**. Also: the model-memory fixes (ONNX arena bound, optional shared model daemon), the prime hook recalls with the prompt's question rather than its paste, and seven lifecycle fixes (PR #7). |
 | **0.17.0** | `AWM_FEEDBACK_NAMES` shipped a **list of 22 real first names** as its built-in default — one deployment's staff list, published in every npm tarball because `src/` is in `files`. It now ships empty, which turns user-feedback auto-promotion off until an operator configures it; **breaking if you relied on the built-in list**, and the migration is one env var. Separately, `memory_write`'s `supersedes` argument was silently discarded whenever reinforce-on-duplicate fired — which is every correction, since a correction restates what it corrects and therefore matches on concept. The engram you asked to retire stayed live, with your correction merged into it. Re-run `awm setup` for the corrected guidance |
 | **0.16.1** | A tokenizer option the reranker had been passing since 0.5.2 turned out never to have existed in transformers.js — silently ignored for six months, and the library default was what the code wanted anyway, so no ranking behavior changes. Found by a dependency sweep that also clears every advisory in the published tree: an install of 0.16.0 audits at 7 (3 high), this one at **0**. Accuracy reproduced identically across two transformers majors, token spend to the token |
 | **0.16.0** | Two recall defaults changed, each after being measured for the first time. A pseudo-relevance feedback pass that ran on **every** query is now opt-in — it cost 7% of every recall and bought nothing measurable across 950 probes. A reranker shortcut that had never once fired on any store benchmarked here is deleted; forced on, it cost success@1 93.0% → 91.7%. Identifier recall is now **93.0%** success@1 / **97.0%** success@5. Re-run `awm setup` for the corrected agent guidance |
@@ -298,18 +301,18 @@ local ONNX via `@huggingface/transformers` — bge-small-en-v1.5 embeddings, ms-
 flan-t5-small expansion. Node 22+.
 
 ```bash
-npx vitest run          # 810 tests
+npx vitest run          # 1071 tests
 npm run bench           # benchmark suites
 npm run build:plugin    # regenerate the Claude Code plugin from src/
 npm run test:docker     # clean-room install of the packed tarball
-npm run test:linux      # build + full suite on Linux (810/810)
+npm run test:linux      # build + full suite on Linux
 ```
 
 ---
 
 ## Status
 
-Active development, v0.17.0. Core retrieval, consolidation, MCP integration, hooks, task
+Active development, v0.19.0. Core retrieval, consolidation, MCP integration, hooks, task
 management, and the HTTP API are stable and in daily production use. PGlite backend stable;
 networked Postgres experimental.
 

package/dist/adapters/common.d.ts

@@ -19,6 +19,40 @@ export declare function isInsidePackage(dbPath: string): boolean;
  * the same product had two different defaults depending on how it was installed.
  */
 export declare function resolveDbPath(packageRoot: string, explicit?: string | null): string;
+/** The default live store, i.e. what a path resolves to when nothing was specified. */
+export declare function isLiveStorePath(dbPath: string): boolean;
+/**
+ * True when this process is a test or benchmark run.
+ *
+ * `VITEST` is set by the runner itself, so it needs no cooperation from the test.
+ * `AWM_BENCH=1` is for a bench or eval script to set; nothing sets it automatically.
+ * Both are absent in every production path: the MCP server, the HTTP server, the CLI
+ * and the hook sidecars.
+ *
+ * Not `NODE_ENV=test`: a developer shell that exports it would hand it to the MCP server
+ * Claude Code starts from there, and the server would refuse its own store and exit,
+ * which shows up only as "Connection closed".
+ */
+export declare function isTestOrBenchContext(): boolean;
+/**
+ * Refuse to open the LIVE store from a test or benchmark process (A-3, 2026-10-12).
+ *
+ * Every launcher defaults `AWM_DB_PATH` to `~/.awm/memory.db` — correct for the product
+ * and wrong for a harness that forgot to override it, because the harness writes.
+ * 4,745 LoCoMo benchmark fixture engrams were found in the live store across 51 unpinned
+ * UUID agent spaces and purged on 2026-10-11; nine real memories sat among them and a
+ * naive filter would have destroyed those too.
+ *
+ * The current eval path is not the offender — `tests/realstore-eval/` reads
+ * `snapshot/store-backfilled.db` and the newest fixture found was dated 2026-06-17, four
+ * months earlier. So this is a guard against the DEFAULT, placed where the default is
+ * produced, not a fix for a live bug.
+ *
+ * Escape hatch: `AWM_ALLOW_LIVE_STORE=1`, for the deliberate case of measuring against
+ * the real store (which `docs/eval-gap-2026-10-11.md` does want, read-only). Loud and
+ * opt-in beats a silent default.
+ */
+export declare function assertNotLiveStoreUnderTest(dbPath: string): void;
 /** Read or generate the hook secret token. */
 export declare function resolveHookSecret(dbPath: string): string;
 /**

package/dist/adapters/common.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"common.d.ts","sourceRoot":"","sources":["../../src/adapters/common.ts"],"names":[],"mappings":"AAeA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAM/C,gEAAgE;AAChE,wBAAgB,kBAAkB,IAAI,MAAM,CAG3C;AAED,sFAAsF;AACtF,wBAAgB,eAAe,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAEvD;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,aAAa,CAAC,WAAW,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,GAAG,MAAM,CAOnF;AAED,8CAA8C;AAC9C,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAUxD;AAED;;;;;;;;GAQG;AACH,eAAO,MAAM,eAAe,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAI5D,CAAC;AAEF,8FAA8F;AAC9F,eAAO,MAAM,cAAc,mHAAoH,CAAC;AAEhJ;;;;;;;GAOG;AACH,wBAAgB,YAAY,CAC1B,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,MAAM,EACf,QAAQ,EAAE,MAAM,EAChB,UAAU,EAAE,MAAM,EAClB,SAAS,EAAE,OAAO,EAClB,IAAI,GAAE;IAAE,aAAa,CAAC,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,IAAI,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAA;CAAO,GAC/F,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CASxB;AAED;;;;;GAKG;AACH,wBAAgB,iBAAiB,CAAC,GAAG,EAAE,YAAY,GAAG;IACpD,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,EAAE,CAAC;CAChB,CAkBA;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE;IACtC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACvB,QAAQ,EAAE,OAAO,CAAC;IAClB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,qFAAqF;IACrF,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,IAAI,CAAC;IAC5C,uFAAuF;IACvF,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB,GAAG,YAAY,CA2Df;AAED;;;GAGG;AACH,wBAAgB,OAAO,IAAI,MAAM,CAEhC;AAID;;;GAGG;AACH;;;;;;;;;;;;;GAaG;AACH;;iEAEiE;AACjE,eAAO,MAAM,aAAa,QAEsD,CAAC;AACjF,eAAO,MAAM,WAAW,+BAA+B,CAAC;AAOxD,wBAAgB,gBAAgB,CAC9B,QAAQ,EAAE,MAAM,EAChB,UAAU,EAAE,MAAM,EAClB,OAAO,GAAE;IAAE,UAAU,CAAC,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IAAC,KAAK,CAAC,EAAE,OAAO,CAAA;CAAO,GACtE,MAAM,CAkFR;AAED,eAAO,MAAM,uBAAuB,QA6bvB,CAAC"}
+{"version":3,"file":"common.d.ts","sourceRoot":"","sources":["../../src/adapters/common.ts"],"names":[],"mappings":"AAeA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAM/C,gEAAgE;AAChE,wBAAgB,kBAAkB,IAAI,MAAM,CAG3C;AAED,sFAAsF;AACtF,wBAAgB,eAAe,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAEvD;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,aAAa,CAAC,WAAW,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,GAAG,MAAM,CAOnF;AAED,uFAAuF;AACvF,wBAAgB,eAAe,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAIvD;AAgBD;;;;;;;;;;;GAWG;AACH,wBAAgB,oBAAoB,IAAI,OAAO,CAG9C;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAgB,2BAA2B,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI,CAYhE;AAED,8CAA8C;AAC9C,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAUxD;AAED;;;;;;;;GAQG;AACH,eAAO,MAAM,eAAe,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAI5D,CAAC;AAEF,8FAA8F;AAC9F,eAAO,MAAM,cAAc,mHAAoH,CAAC;AAEhJ;;;;;;;GAOG;AACH,wBAAgB,YAAY,CAC1B,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,MAAM,EACf,QAAQ,EAAE,MAAM,EAChB,UAAU,EAAE,MAAM,EAClB,SAAS,EAAE,OAAO,EAClB,IAAI,GAAE;IAAE,aAAa,CAAC,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,IAAI,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAA;CAAO,GAC/F,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CASxB;AAED;;;;;GAKG;AACH,wBAAgB,iBAAiB,CAAC,GAAG,EAAE,YAAY,GAAG;IACpD,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,EAAE,CAAC;CAChB,CAkBA;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE;IACtC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACvB,QAAQ,EAAE,OAAO,CAAC;IAClB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,qFAAqF;IACrF,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,IAAI,CAAC;IAC5C,uFAAuF;IACvF,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB,GAAG,YAAY,CA2Df;AAED;;;GAGG;AACH,wBAAgB,OAAO,IAAI,MAAM,CAEhC;AAID;;;GAGG;AACH;;;;;;;;;;;;;GAaG;AACH;;iEAEiE;AACjE,eAAO,MAAM,aAAa,QAEsD,CAAC;AACjF,eAAO,MAAM,WAAW,+BAA+B,CAAC;AAOxD,wBAAgB,gBAAgB,CAC9B,QAAQ,EAAE,MAAM,EAChB,UAAU,EAAE,MAAM,EAClB,OAAO,GAAE;IAAE,UAAU,CAAC,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IAAC,KAAK,CAAC,EAAE,OAAO,CAAA;CAAO,GACtE,MAAM,CAkFR;AAED,eAAO,MAAM,uBAAuB,QAscvB,CAAC"}