agent-working-memory 0.16.0 → 0.16.1

package/README.md

@@ -181,10 +181,11 @@ compatible, and the store is the same SQLite file before and after.
 
 ---
 
-## What's new — v0.16.0
+## What's new — v0.16.1
 
 | | |
 |---|---|
+| **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 |
 | **0.15.9** | A memory that consolidation archived as a near-duplicate — even at high confidence — could never be reinforced again, so the same recurring fact kept spawning new copies instead of strengthening one. An archived match is now revived on reinforcement |
 | **0.15.6–8** | **Data-loss fix.** The default store used to sit inside the installed npm package, where `npm install -g` deletes it on upgrade. It is now `~/.awm/memory.db`; re-running `awm setup` copies an affected store to safety and `awm doctor` fails on it. 0.15.7 corrects the docs that still taught the old location |
@@ -307,7 +308,7 @@ npm run test:linux      # build + full suite on Linux (802/802)
 
 ## Status
 
-Active development, v0.16.0. Core retrieval, consolidation, MCP integration, hooks, task
+Active development, v0.16.1. 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/core/reranker.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"reranker.d.ts","sourceRoot":"","sources":["../../src/core/reranker.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;GAWG;AA6GH,iCAAiC;AACjC,wBAAsB,WAAW,IAAI,OAAO,CAAC,GAAG,CAAC,CAGhD;AAED,MAAM,WAAW,YAAY;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;CACf;AAED;;;;GAIG;AACH,wBAAsB,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,YAAY,EAAE,CAAC,CAGvF"}
+{"version":3,"file":"reranker.d.ts","sourceRoot":"","sources":["../../src/core/reranker.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;GAWG;AAqHH,iCAAiC;AACjC,wBAAsB,WAAW,IAAI,OAAO,CAAC,GAAG,CAAC,CAGhD;AAED,MAAM,WAAW,YAAY;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;CACf;AAED;;;;GAIG;AACH,wBAAsB,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,YAAY,EAAE,CAAC,CAGvF"}

package/dist/core/reranker.js

@@ -68,7 +68,14 @@ async function inProcessRerank(args) {
             text_pair: passages,
             padding: true,
             truncation: true,
-            return_tensors: 'pt',
+            // NOT 'return_tensors'. That key never existed in transformers.js -- it is the
+            // PyTorch Python spelling -- so from 0.5.2 (commit 1ccbf95, 2026-03-21) until
+            // 2026-10-09 it was silently ignored and the library default
+            // (return_tensor=true, i.e. Tensors) applied.
+            // Stating it explicitly preserves exactly the behaviour that was already
+            // running, and survives a future change of default. transformers 4.x turned
+            // the typo into a type error, which is the only reason it was ever noticed.
+            return_tensor: true,
         });
         const output = await model(inputs);
         const logits = output.logits ?? output.last_hidden_state;
@@ -90,7 +97,8 @@ async function inProcessRerank(args) {
                     text_pair: passages[i],
                     padding: true,
                     truncation: true,
-                    return_tensors: 'pt',
+                    // Same ignored-key story as the batch path above.
+                    return_tensor: true,
                 });
                 const output = await model(inputs);
                 const logits = output.logits ?? output.last_hidden_state;

package/dist/core/reranker.js.map

@@ -1 +1 @@
-{"version":3,"file":"reranker.js","sourceRoot":"","sources":["../../src/core/reranker.ts"],"names":[],"mappings":"AAAA,gDAAgD;AAChD,sCAAsC;AACtC;;;;;;;;;;;GAWG;AAEH,OAAO,EACL,aAAa,EACb,kCAAkC,GAGnC,MAAM,2BAA2B,CAAC;AACnC,OAAO,EAAE,cAAc,EAAE,yBAAyB,EAAE,MAAM,gBAAgB,CAAC;AAC3E,OAAO,EAAE,mBAAmB,EAAE,MAAM,kBAAkB,CAAC;AAEvD,MAAM,aAAa,GAAG,+BAA+B,CAAC;AACtD,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,kBAAkB,IAAI,aAAa,CAAC;AAEjE;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,KAAK,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,kBAAkB,IAAI,MAAM,CAAqD,CAAC;AAE7G,8BAA8B;AAE9B,IAAI,SAAS,GAA+B,IAAI,CAAC;AACjD,IAAI,KAAK,GAA2B,IAAI,CAAC;AACzC,IAAI,WAAW,GAAyB,IAAI,CAAC;AAE7C,KAAK,UAAU,YAAY;IACzB,IAAI,SAAS,IAAI,KAAK;QAAE,OAAO;IAC/B,IAAI,WAAW;QAAE,OAAO,WAAW,CAAC;IACpC,WAAW,GAAG,CAAC,KAAK,IAAI,EAAE;QACxB,mBAAmB,EAAE,CAAC;QACtB,SAAS,GAAG,MAAM,aAAa,CAAC,eAAe,CAAC,QAAQ,CAAC,CAAC;QAC1D,KAAK,GAAG,MAAM,kCAAkC,CAAC,eAAe,CAAC,QAAQ,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,CAAC;QAC7F,OAAO,CAAC,KAAK,CAAC,sCAAsC,QAAQ,KAAK,KAAK,GAAG,CAAC,CAAC;IAC7E,CAAC,CAAC,EAAE,CAAC;IACL,OAAO,WAAW,CAAC;AACrB,CAAC;AAED,SAAS,OAAO,CAAC,CAAS;IACxB,OAAO,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AAChC,CAAC;AAED,KAAK,UAAU,eAAe,CAAC,IAA2C;IACxE,MAAM,EAAE,KAAK,EAAE,QAAQ,EAAE,GAAG,IAAI,CAAC;IACjC,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IACrC,MAAM,YAAY,EAAE,CAAC;IAErB,aAAa;IACb,IAAI,CAAC;QACH,MAAM,OAAO,GAAG,QAAQ,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC,CAAC;QAC1C,MAAM,MAAM,GAAG,SAAU,CAAC,OAAO,EAAE;YACjC,SAAS,EAAE,QAAQ;YACnB,OAAO,EAAE,IAAI;YACb,UAAU,EAAE,IAAI;YAChB,cAAc,EAAE,IAAI;SACrB,CAAC,CAAC;QACH,MAAM,MAAM,GAAG,MAAM,KAAM,CAAC,MAAM,CAAC,CAAC;QACpC,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,MAAM,CAAC,iBAAiB,CAAC;QACzD,MAAM,IAAI,GAAG,MAAM,CAAC,IAA+B,CAAC;QACpD,MAAM,OAAO,GAA4C,EAAE,CAAC;QAC5D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACzC,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;YACtC,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;QACvD,CAAC;QACD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;QAC1C,OAAO,OAAO,CAAC;IACjB,CAAC;IAAC,MAAM,CAAC;QACP,kDAAkD;QAClD,MAAM,OAAO,GAA4C,EAAE,CAAC;QAC5D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACzC,IAAI,CAAC;gBACH,MAAM,MAAM,GAAG,SAAU,CAAC,KAAK,EAAE;oBAC/B,SAAS,EAAE,QAAQ,CAAC,CAAC,CAAC;oBACtB,OAAO,EAAE,IAAI;oBACb,UAAU,EAAE,IAAI;oBAChB,cAAc,EAAE,IAAI;iBACrB,CAAC,CAAC;gBACH,MAAM,MAAM,GAAG,MAAM,KAAM,CAAC,MAAM,CAAC,CAAC;gBACpC,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,MAAM,CAAC,iBAAiB,CAAC;gBACzD,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAW,CAAC;gBAC1C,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;YACvD,CAAC;YAAC,MAAM,CAAC;gBACP,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC;YACvC,CAAC;QACH,CAAC;QACD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;QAC1C,OAAO,OAAO,CAAC;IACjB,CAAC;AACH,CAAC;AAED,gDAAgD;AAChD,yBAAyB,CAAC,EAAE,MAAM,EAAE,eAAe,EAAE,CAAC,CAAC;AAEvD,qBAAqB;AAErB,iCAAiC;AACjC,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,MAAM,YAAY,EAAE,CAAC;IACrB,OAAO,KAAK,CAAC;AACf,CAAC;AAOD;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,MAAM,CAAC,KAAa,EAAE,QAAkB;IAC5D,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IACrC,OAAO,cAAc,CAAC,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;AAC7C,CAAC"}
+{"version":3,"file":"reranker.js","sourceRoot":"","sources":["../../src/core/reranker.ts"],"names":[],"mappings":"AAAA,gDAAgD;AAChD,sCAAsC;AACtC;;;;;;;;;;;GAWG;AAEH,OAAO,EACL,aAAa,EACb,kCAAkC,GAGnC,MAAM,2BAA2B,CAAC;AACnC,OAAO,EAAE,cAAc,EAAE,yBAAyB,EAAE,MAAM,gBAAgB,CAAC;AAC3E,OAAO,EAAE,mBAAmB,EAAE,MAAM,kBAAkB,CAAC;AAEvD,MAAM,aAAa,GAAG,+BAA+B,CAAC;AACtD,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,kBAAkB,IAAI,aAAa,CAAC;AAEjE;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,KAAK,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,kBAAkB,IAAI,MAAM,CAAqD,CAAC;AAE7G,8BAA8B;AAE9B,IAAI,SAAS,GAA+B,IAAI,CAAC;AACjD,IAAI,KAAK,GAA2B,IAAI,CAAC;AACzC,IAAI,WAAW,GAAyB,IAAI,CAAC;AAE7C,KAAK,UAAU,YAAY;IACzB,IAAI,SAAS,IAAI,KAAK;QAAE,OAAO;IAC/B,IAAI,WAAW;QAAE,OAAO,WAAW,CAAC;IACpC,WAAW,GAAG,CAAC,KAAK,IAAI,EAAE;QACxB,mBAAmB,EAAE,CAAC;QACtB,SAAS,GAAG,MAAM,aAAa,CAAC,eAAe,CAAC,QAAQ,CAAC,CAAC;QAC1D,KAAK,GAAG,MAAM,kCAAkC,CAAC,eAAe,CAAC,QAAQ,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,CAAC;QAC7F,OAAO,CAAC,KAAK,CAAC,sCAAsC,QAAQ,KAAK,KAAK,GAAG,CAAC,CAAC;IAC7E,CAAC,CAAC,EAAE,CAAC;IACL,OAAO,WAAW,CAAC;AACrB,CAAC;AAED,SAAS,OAAO,CAAC,CAAS;IACxB,OAAO,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AAChC,CAAC;AAED,KAAK,UAAU,eAAe,CAAC,IAA2C;IACxE,MAAM,EAAE,KAAK,EAAE,QAAQ,EAAE,GAAG,IAAI,CAAC;IACjC,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IACrC,MAAM,YAAY,EAAE,CAAC;IAErB,aAAa;IACb,IAAI,CAAC;QACH,MAAM,OAAO,GAAG,QAAQ,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC,CAAC;QAC1C,MAAM,MAAM,GAAG,SAAU,CAAC,OAAO,EAAE;YACjC,SAAS,EAAE,QAAQ;YACnB,OAAO,EAAE,IAAI;YACb,UAAU,EAAE,IAAI;YAChB,+EAA+E;YAC/E,8EAA8E;YAC9E,6DAA6D;YAC7D,8CAA8C;YAC9C,yEAAyE;YACzE,4EAA4E;YAC5E,4EAA4E;YAC5E,aAAa,EAAE,IAAI;SACpB,CAAC,CAAC;QACH,MAAM,MAAM,GAAG,MAAM,KAAM,CAAC,MAAM,CAAC,CAAC;QACpC,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,MAAM,CAAC,iBAAiB,CAAC;QACzD,MAAM,IAAI,GAAG,MAAM,CAAC,IAA+B,CAAC;QACpD,MAAM,OAAO,GAA4C,EAAE,CAAC;QAC5D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACzC,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;YACtC,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;QACvD,CAAC;QACD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;QAC1C,OAAO,OAAO,CAAC;IACjB,CAAC;IAAC,MAAM,CAAC;QACP,kDAAkD;QAClD,MAAM,OAAO,GAA4C,EAAE,CAAC;QAC5D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACzC,IAAI,CAAC;gBACH,MAAM,MAAM,GAAG,SAAU,CAAC,KAAK,EAAE;oBAC/B,SAAS,EAAE,QAAQ,CAAC,CAAC,CAAC;oBACtB,OAAO,EAAE,IAAI;oBACb,UAAU,EAAE,IAAI;oBAChB,kDAAkD;oBAClD,aAAa,EAAE,IAAI;iBACpB,CAAC,CAAC;gBACH,MAAM,MAAM,GAAG,MAAM,KAAM,CAAC,MAAM,CAAC,CAAC;gBACpC,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,MAAM,CAAC,iBAAiB,CAAC;gBACzD,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAW,CAAC;gBAC1C,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;YACvD,CAAC;YAAC,MAAM,CAAC;gBACP,OAAO,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC;YACvC,CAAC;QACH,CAAC;QACD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;QAC1C,OAAO,OAAO,CAAC;IACjB,CAAC;AACH,CAAC;AAED,gDAAgD;AAChD,yBAAyB,CAAC,EAAE,MAAM,EAAE,eAAe,EAAE,CAAC,CAAC;AAEvD,qBAAqB;AAErB,iCAAiC;AACjC,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,MAAM,YAAY,EAAE,CAAC;IACrB,OAAO,KAAK,CAAC;AACf,CAAC;AAOD;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,MAAM,CAAC,KAAa,EAAE,QAAkB;IAC5D,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IACrC,OAAO,cAAc,CAAC,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;AAC7C,CAAC"}

package/package.json

@@ -1,106 +1,106 @@
-{
-  "name": "agent-working-memory",
-  "version": "0.16.0",
-  "mcpName": "io.github.CompleteIdeas/agent-working-memory",
-  "description": "Persistent, local-first memory for AI agents. Selective by design: filters what's worth keeping, links what's related, and says nothing when it doesn't know. SQLite + ONNX, no cloud, no API keys.",
-  "keywords": [
-    "ai-agents",
-    "agent-memory",
-    "memory",
-    "mcp",
-    "model-context-protocol",
-    "claude-code",
-    "local-first",
-    "sqlite",
-    "onnx",
-    "retrieval",
-    "cognitive-architecture",
-    "act-r",
-    "hebbian",
-    "embeddings",
-    "reranker",
-    "typescript"
-  ],
-  "type": "module",
-  "main": "dist/index.js",
-  "bin": {
-    "awm": "dist/cli.js"
-  },
-  "scripts": {
-    "build": "tsc",
-    "build:plugin": "node scripts/build-plugin.mjs",
-    "build:mcpb": "node scripts/build-mcpb.mjs",
-    "pack:mcpb": "node scripts/build-mcpb.mjs --pack",
-    "prepublishOnly": "npm run build && npm run build:plugin && npm run build:mcpb && npm run check:release",
-    "dev": "tsx watch src/index.ts",
-    "start": "node dist/index.js",
-    "test": "vitest",
-    "test:run": "vitest run",
-    "mcp": "tsx src/mcp.ts",
-    "test:mcp": "tsx tests/mcp-smoke.ts",
-    "test:self": "tsx tests/self-test/runner.ts",
-    "test:workday": "tsx tests/workday-eval/runner.ts",
-    "test:locomo": "tsx tests/locomo-eval/runner.ts",
-    "test:tokens": "tsx tests/token-savings/runner.ts",
-    "test:realworld": "tsx tests/realworld-eval/runner.ts",
-    "test:ab": "tsx tests/ab-test/runner.ts",
-    "test:sleep": "tsx tests/sleep-cycle/runner.ts",
-    "test:stress": "tsx tests/stress-test/runner.ts",
-    "test:edge": "tsx tests/edge-cases/runner.ts",
-    "test:pilot": "tsx tests/pilot-metrics/runner.ts",
-    "test:perf": "tsx tests/perf-test/runner.ts",
-    "eval": "tsx tests/eval/run.ts",
-    "eval:generate": "tsx tests/eval/generate.ts",
-    "test:all": "vitest run && tsx tests/mcp-smoke.ts",
-    "check:release": "node scripts/check-release.mjs",
-    "bench": "node scripts/bench-current.mjs",
-    "bench:public:build": "tsx tests/realstore-eval/build-public-snapshot.ts",
-    "bench:public": "node scripts/bench-public.mjs",
-    "profile:recall": "node scripts/profile-recall.mjs",
-    "test:docker": "bash scripts/docker-release-test.sh",
-    "test:linux": "bash scripts/docker-release-test.sh linux",
-    "test:desktop": "bash scripts/docker-release-test.sh desktop"
-  },
-  "engines": {
-    "node": ">=22.0.0"
-  },
-  "license": "Apache-2.0",
-  "files": [
-    "dist/",
-    "src/",
-    "plugin/",
-    ".claude-plugin/",
-    "package.json",
-    "README.md",
-    "AGENTS.md",
-    "llms.txt",
-    "LICENSE"
-  ],
-  "dependencies": {
-    "@electric-sql/pglite": "^0.5.6",
-    "@electric-sql/pglite-pgvector": "^0.0.7",
-    "@huggingface/transformers": "^3.8.1",
-    "@modelcontextprotocol/sdk": "^1.27.1",
-    "@toon-format/toon": "^2.3.0",
-    "better-sqlite3": "^12.6.2",
-    "fastify": "^5.8.2",
-    "pg": "^8.13.1",
-    "zod": "^4.3.6"
-  },
-  "devDependencies": {
-    "@types/better-sqlite3": "^7.6.13",
-    "@types/node": "^25.3.5",
-    "@types/pg": "^8.11.10",
-    "tsx": "^4.21.0",
-    "typescript": "^5.9.3",
-    "vitest": "^4.0.18"
-  },
-  "repository": {
-    "type": "git",
-    "url": "git+https://github.com/CompleteIdeas/agent-working-memory.git"
-  },
-  "homepage": "https://github.com/CompleteIdeas/agent-working-memory#readme",
-  "bugs": {
-    "url": "https://github.com/CompleteIdeas/agent-working-memory/issues"
-  }
-}
+{
+  "name": "agent-working-memory",
+  "version": "0.16.1",
+  "mcpName": "io.github.CompleteIdeas/agent-working-memory",
+  "description": "Persistent, local-first memory for AI agents. Selective by design: filters what's worth keeping, links what's related, and says nothing when it doesn't know. SQLite + ONNX, no cloud, no API keys.",
+  "keywords": [
+    "ai-agents",
+    "agent-memory",
+    "memory",
+    "mcp",
+    "model-context-protocol",
+    "claude-code",
+    "local-first",
+    "sqlite",
+    "onnx",
+    "retrieval",
+    "cognitive-architecture",
+    "act-r",
+    "hebbian",
+    "embeddings",
+    "reranker",
+    "typescript"
+  ],
+  "type": "module",
+  "main": "dist/index.js",
+  "bin": {
+    "awm": "dist/cli.js"
+  },
+  "scripts": {
+    "build": "tsc",
+    "build:plugin": "node scripts/build-plugin.mjs",
+    "build:mcpb": "node scripts/build-mcpb.mjs",
+    "pack:mcpb": "node scripts/build-mcpb.mjs --pack",
+    "prepublishOnly": "npm run build && npm run build:plugin && npm run build:mcpb && npm run check:release",
+    "dev": "tsx watch src/index.ts",
+    "start": "node dist/index.js",
+    "test": "vitest",
+    "test:run": "vitest run",
+    "mcp": "tsx src/mcp.ts",
+    "test:mcp": "tsx tests/mcp-smoke.ts",
+    "test:self": "tsx tests/self-test/runner.ts",
+    "test:workday": "tsx tests/workday-eval/runner.ts",
+    "test:locomo": "tsx tests/locomo-eval/runner.ts",
+    "test:tokens": "tsx tests/token-savings/runner.ts",
+    "test:realworld": "tsx tests/realworld-eval/runner.ts",
+    "test:ab": "tsx tests/ab-test/runner.ts",
+    "test:sleep": "tsx tests/sleep-cycle/runner.ts",
+    "test:stress": "tsx tests/stress-test/runner.ts",
+    "test:edge": "tsx tests/edge-cases/runner.ts",
+    "test:pilot": "tsx tests/pilot-metrics/runner.ts",
+    "test:perf": "tsx tests/perf-test/runner.ts",
+    "eval": "tsx tests/eval/run.ts",
+    "eval:generate": "tsx tests/eval/generate.ts",
+    "test:all": "vitest run && tsx tests/mcp-smoke.ts",
+    "check:release": "node scripts/check-release.mjs",
+    "bench": "node scripts/bench-current.mjs",
+    "bench:public:build": "tsx tests/realstore-eval/build-public-snapshot.ts",
+    "bench:public": "node scripts/bench-public.mjs",
+    "profile:recall": "node scripts/profile-recall.mjs",
+    "test:docker": "bash scripts/docker-release-test.sh",
+    "test:linux": "bash scripts/docker-release-test.sh linux",
+    "test:desktop": "bash scripts/docker-release-test.sh desktop"
+  },
+  "engines": {
+    "node": ">=22.0.0"
+  },
+  "license": "Apache-2.0",
+  "files": [
+    "dist/",
+    "src/",
+    "plugin/",
+    ".claude-plugin/",
+    "package.json",
+    "README.md",
+    "AGENTS.md",
+    "llms.txt",
+    "LICENSE"
+  ],
+  "dependencies": {
+    "@electric-sql/pglite": "0.5.6",
+    "@electric-sql/pglite-pgvector": "^0.0.7",
+    "@huggingface/transformers": "^4.3.1",
+    "@modelcontextprotocol/sdk": "^1.32.1",
+    "@toon-format/toon": "^2.3.0",
+    "better-sqlite3": "^12.6.2",
+    "fastify": "^5.12.5",
+    "pg": "^8.13.1",
+    "zod": "^4.3.6"
+  },
+  "devDependencies": {
+    "@types/better-sqlite3": "^7.6.13",
+    "@types/node": "^25.3.5",
+    "@types/pg": "^8.11.10",
+    "tsx": "^4.21.0",
+    "typescript": "^5.9.3",
+    "vitest": "^4.0.18"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/CompleteIdeas/agent-working-memory.git"
+  },
+  "homepage": "https://github.com/CompleteIdeas/agent-working-memory#readme",
+  "bugs": {
+    "url": "https://github.com/CompleteIdeas/agent-working-memory/issues"
+  }
+}

package/plugin/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "awm",
   "displayName": "Agent Working Memory",
-  "version": "0.16.0",
+  "version": "0.16.1",
   "description": "Persistent, local-first memory for AI agents. Selective by design: filters what's worth keeping, links what's related, and says nothing when it doesn't know. SQLite + ONNX, no cloud, no API keys.",
   "author": {
     "name": "Robert Winter"

package/plugin/README.md

@@ -22,7 +22,7 @@ The launcher defaults `AWM_DB_PATH` to `~/.awm/memory.db` — **the same file
 than quietly starting an empty one. Set `AWM_DB_PATH` yourself to override.
 
 It prefers an installed copy of the package (global, or a local `node_modules`) and falls
-back to `npx -p agent-working-memory@0.16.0`. The npx path rebuilds
+back to `npx -p agent-working-memory@0.16.1`. The npx path rebuilds
 `better-sqlite3` on a cold cache, which is slow enough to look like a hang, so a real
 install is worth having:
 

package/plugin/bin/awm-mcp-launcher.cjs

@@ -6,7 +6,7 @@
 //
 // This is a REAL source file, not a string inside the generator, because it is real code:
 // it gets syntax-checked, linted and tested like everything else. `npm run build:plugin`
-// copies it to plugin/bin/ and substitutes 0.16.0.
+// copies it to plugin/bin/ and substitutes 0.16.1.
 //
 // Two jobs a bare `mcpServers` entry in plugin.json cannot do:
 //
@@ -95,7 +95,7 @@ if (found) {
   const isWin = process.platform === 'win32';
   const child = spawn(
     isWin ? 'npx.cmd' : 'npx',
-    ['-y', '-p', 'agent-working-memory@0.16.0', 'awm', 'mcp'],
+    ['-y', '-p', 'agent-working-memory@0.16.1', 'awm', 'mcp'],
     { stdio: 'inherit', env: process.env, shell: isWin }
   );
   child.on('error', (e) => {

package/plugin/skills/awm-memory/SKILL.md

@@ -10,7 +10,7 @@ description: >-
 
 # Agent Working Memory
 
-Generated from `src/adapters/common.ts` (AWM v0.16.0) — the same guidance
+Generated from `src/adapters/common.ts` (AWM v0.16.1) — the same guidance
 `awm setup` writes into CLAUDE.md. Do not edit here; edit the source and re-run
 `npm run build:plugin`.
 

package/src/core/reranker.ts

@@ -1,142 +1,150 @@
-// Copyright 2026 Robert Winter / Complete Ideas
-// SPDX-License-Identifier: Apache-2.0
-/**
- * Cross-Encoder Re-Ranker - scores (query, passage) pairs for relevance.
- *
- * Uses Xenova/ms-marco-MiniLM-L-6-v2 (91MB ONNX at the fp32 default; the 23MB
- * quantized file is what AWM_RERANKER_DTYPE=q8 loads) trained on MS-MARCO
- * passage ranking. Unlike bi-encoders, cross-encoders see both query and
- * passage together via full attention - much better at judging if a
- * passage actually answers a question.
- *
- * AWM 0.8.x: inference dispatches through ml-worker.ts (currently in-process
- * — see ml-worker.ts for the worker_threads → in-process revert rationale).
- */
-
-import {
-  AutoTokenizer,
-  AutoModelForSequenceClassification,
-  type PreTrainedTokenizer,
-  type PreTrainedModel,
-} from '@huggingface/transformers';
-import { dispatchRerank, registerInProcessHandlers } from './ml-worker.js';
-import { ensureModelCacheDir } from './model-cache.js';
-
-const DEFAULT_MODEL = 'Xenova/ms-marco-MiniLM-L-6-v2';
-const MODEL_ID = process.env.AWM_RERANKER_MODEL || DEFAULT_MODEL;
-
-/**
- * Weight precision for the cross-encoder. fp32 is the shipped default and the
- * only value any published number was measured at.
- *
- * It is exposed because the cross-encoder is 83% of warm recall latency
- * (measured, `npm run profile:recall`), which makes precision the largest
- * single latency lever in the system — and an unmeasurable one while it was a
- * literal. RE-MEASURED 2026-10-09 on the pipeline that ships: 'q8' cuts total
- * recall latency 26-33% across three suites, for ONE QUERY in 450 — identifier
- * success@1 unchanged at 93.0%, topic 92.2% -> 92.0%. Correct abstention is
- * unmoved on all three suites; success@5 is unmoved on both private suites and
- * moves +0.5pp on the public corpus, inside its 4pp seed-noise floor. It is deterministic (two runs agree
- * on all 450 queries to 1e-6), and because no model is bundled it SHRINKS the
- * first-run fetch: the ONNX file it loads is 23MB against fp32's 91MB. 'q4' is dominated: equal
- * accuracy, only -3% latency, 6.1s cold start against fp32's 1.6s. The default
- * is untouched because it is a product decision, not because the cost is
- * unknown; see docs/recall-latency.md.
- */
-const DTYPE = (process.env.AWM_RERANKER_DTYPE || 'fp32') as 'fp32' | 'fp16' | 'q8' | 'int8' | 'uint8' | 'q4';
-
-// --- In-process fallback ---
-
-let tokenizer: PreTrainedTokenizer | null = null;
-let model: PreTrainedModel | null = null;
-let initPromise: Promise<void> | null = null;
-
-async function ensureLoaded(): Promise<void> {
-  if (tokenizer && model) return;
-  if (initPromise) return initPromise;
-  initPromise = (async () => {
-    ensureModelCacheDir();
-    tokenizer = await AutoTokenizer.from_pretrained(MODEL_ID);
-    model = await AutoModelForSequenceClassification.from_pretrained(MODEL_ID, { dtype: DTYPE });
-    console.error(`Re-ranker model loaded in-process: ${MODEL_ID} (${DTYPE})`);
-  })();
-  return initPromise;
-}
-
-function sigmoid(x: number): number {
-  return 1 / (1 + Math.exp(-x));
-}
-
-async function inProcessRerank(args: { query: string; passages: string[] }): Promise<Array<{ index: number; score: number }>> {
-  const { query, passages } = args;
-  if (passages.length === 0) return [];
-  await ensureLoaded();
-
-  // Batch path
-  try {
-    const queries = passages.map(() => query);
-    const inputs = tokenizer!(queries, {
-      text_pair: passages,
-      padding: true,
-      truncation: true,
-      return_tensors: 'pt',
-    });
-    const output = await model!(inputs);
-    const logits = output.logits ?? output.last_hidden_state;
-    const data = logits.data as Float32Array | number[];
-    const results: Array<{ index: number; score: number }> = [];
-    for (let i = 0; i < passages.length; i++) {
-      const rawLogit = Number(data[i] ?? 0);
-      results.push({ index: i, score: sigmoid(rawLogit) });
-    }
-    results.sort((a, b) => b.score - a.score);
-    return results;
-  } catch {
-    // Per-passage fallback (the original 0.7.13 path)
-    const results: Array<{ index: number; score: number }> = [];
-    for (let i = 0; i < passages.length; i++) {
-      try {
-        const inputs = tokenizer!(query, {
-          text_pair: passages[i],
-          padding: true,
-          truncation: true,
-          return_tensors: 'pt',
-        });
-        const output = await model!(inputs);
-        const logits = output.logits ?? output.last_hidden_state;
-        const rawLogit = logits.data[0] as number;
-        results.push({ index: i, score: sigmoid(rawLogit) });
-      } catch {
-        results.push({ index: i, score: 0 });
-      }
-    }
-    results.sort((a, b) => b.score - a.score);
-    return results;
-  }
-}
-
-// Register the in-process handler with the pool
-registerInProcessHandlers({ rerank: inProcessRerank });
-
-// --- Public API ---
-
-/** Kept for backwards compat. */
-export async function getReranker(): Promise<any> {
-  await ensureLoaded();
-  return model;
-}
-
-export interface RerankResult {
-  index: number;
-  score: number; // sigmoid-normalized relevance (0-1)
-}
-
-/**
- * Re-rank candidate passages against a query using the cross-encoder.
- * Returns results sorted by relevance score (descending).
- * Dispatches to the worker pool (or in-process fallback).
- */
-export async function rerank(query: string, passages: string[]): Promise<RerankResult[]> {
-  if (passages.length === 0) return [];
-  return dispatchRerank({ query, passages });
-}
+// Copyright 2026 Robert Winter / Complete Ideas
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Cross-Encoder Re-Ranker - scores (query, passage) pairs for relevance.
+ *
+ * Uses Xenova/ms-marco-MiniLM-L-6-v2 (91MB ONNX at the fp32 default; the 23MB
+ * quantized file is what AWM_RERANKER_DTYPE=q8 loads) trained on MS-MARCO
+ * passage ranking. Unlike bi-encoders, cross-encoders see both query and
+ * passage together via full attention - much better at judging if a
+ * passage actually answers a question.
+ *
+ * AWM 0.8.x: inference dispatches through ml-worker.ts (currently in-process
+ * — see ml-worker.ts for the worker_threads → in-process revert rationale).
+ */
+
+import {
+  AutoTokenizer,
+  AutoModelForSequenceClassification,
+  type PreTrainedTokenizer,
+  type PreTrainedModel,
+} from '@huggingface/transformers';
+import { dispatchRerank, registerInProcessHandlers } from './ml-worker.js';
+import { ensureModelCacheDir } from './model-cache.js';
+
+const DEFAULT_MODEL = 'Xenova/ms-marco-MiniLM-L-6-v2';
+const MODEL_ID = process.env.AWM_RERANKER_MODEL || DEFAULT_MODEL;
+
+/**
+ * Weight precision for the cross-encoder. fp32 is the shipped default and the
+ * only value any published number was measured at.
+ *
+ * It is exposed because the cross-encoder is 83% of warm recall latency
+ * (measured, `npm run profile:recall`), which makes precision the largest
+ * single latency lever in the system — and an unmeasurable one while it was a
+ * literal. RE-MEASURED 2026-10-09 on the pipeline that ships: 'q8' cuts total
+ * recall latency 26-33% across three suites, for ONE QUERY in 450 — identifier
+ * success@1 unchanged at 93.0%, topic 92.2% -> 92.0%. Correct abstention is
+ * unmoved on all three suites; success@5 is unmoved on both private suites and
+ * moves +0.5pp on the public corpus, inside its 4pp seed-noise floor. It is deterministic (two runs agree
+ * on all 450 queries to 1e-6), and because no model is bundled it SHRINKS the
+ * first-run fetch: the ONNX file it loads is 23MB against fp32's 91MB. 'q4' is dominated: equal
+ * accuracy, only -3% latency, 6.1s cold start against fp32's 1.6s. The default
+ * is untouched because it is a product decision, not because the cost is
+ * unknown; see docs/recall-latency.md.
+ */
+const DTYPE = (process.env.AWM_RERANKER_DTYPE || 'fp32') as 'fp32' | 'fp16' | 'q8' | 'int8' | 'uint8' | 'q4';
+
+// --- In-process fallback ---
+
+let tokenizer: PreTrainedTokenizer | null = null;
+let model: PreTrainedModel | null = null;
+let initPromise: Promise<void> | null = null;
+
+async function ensureLoaded(): Promise<void> {
+  if (tokenizer && model) return;
+  if (initPromise) return initPromise;
+  initPromise = (async () => {
+    ensureModelCacheDir();
+    tokenizer = await AutoTokenizer.from_pretrained(MODEL_ID);
+    model = await AutoModelForSequenceClassification.from_pretrained(MODEL_ID, { dtype: DTYPE });
+    console.error(`Re-ranker model loaded in-process: ${MODEL_ID} (${DTYPE})`);
+  })();
+  return initPromise;
+}
+
+function sigmoid(x: number): number {
+  return 1 / (1 + Math.exp(-x));
+}
+
+async function inProcessRerank(args: { query: string; passages: string[] }): Promise<Array<{ index: number; score: number }>> {
+  const { query, passages } = args;
+  if (passages.length === 0) return [];
+  await ensureLoaded();
+
+  // Batch path
+  try {
+    const queries = passages.map(() => query);
+    const inputs = tokenizer!(queries, {
+      text_pair: passages,
+      padding: true,
+      truncation: true,
+      // NOT 'return_tensors'. That key never existed in transformers.js -- it is the
+      // PyTorch Python spelling -- so from 0.5.2 (commit 1ccbf95, 2026-03-21) until
+      // 2026-10-09 it was silently ignored and the library default
+      // (return_tensor=true, i.e. Tensors) applied.
+      // Stating it explicitly preserves exactly the behaviour that was already
+      // running, and survives a future change of default. transformers 4.x turned
+      // the typo into a type error, which is the only reason it was ever noticed.
+      return_tensor: true,
+    });
+    const output = await model!(inputs);
+    const logits = output.logits ?? output.last_hidden_state;
+    const data = logits.data as Float32Array | number[];
+    const results: Array<{ index: number; score: number }> = [];
+    for (let i = 0; i < passages.length; i++) {
+      const rawLogit = Number(data[i] ?? 0);
+      results.push({ index: i, score: sigmoid(rawLogit) });
+    }
+    results.sort((a, b) => b.score - a.score);
+    return results;
+  } catch {
+    // Per-passage fallback (the original 0.7.13 path)
+    const results: Array<{ index: number; score: number }> = [];
+    for (let i = 0; i < passages.length; i++) {
+      try {
+        const inputs = tokenizer!(query, {
+          text_pair: passages[i],
+          padding: true,
+          truncation: true,
+          // Same ignored-key story as the batch path above.
+          return_tensor: true,
+        });
+        const output = await model!(inputs);
+        const logits = output.logits ?? output.last_hidden_state;
+        const rawLogit = logits.data[0] as number;
+        results.push({ index: i, score: sigmoid(rawLogit) });
+      } catch {
+        results.push({ index: i, score: 0 });
+      }
+    }
+    results.sort((a, b) => b.score - a.score);
+    return results;
+  }
+}
+
+// Register the in-process handler with the pool
+registerInProcessHandlers({ rerank: inProcessRerank });
+
+// --- Public API ---
+
+/** Kept for backwards compat. */
+export async function getReranker(): Promise<any> {
+  await ensureLoaded();
+  return model;
+}
+
+export interface RerankResult {
+  index: number;
+  score: number; // sigmoid-normalized relevance (0-1)
+}
+
+/**
+ * Re-rank candidate passages against a query using the cross-encoder.
+ * Returns results sorted by relevance score (descending).
+ * Dispatches to the worker pool (or in-process fallback).
+ */
+export async function rerank(query: string, passages: string[]): Promise<RerankResult[]> {
+  if (passages.length === 0) return [];
+  return dispatchRerank({ query, passages });
+}