agenr 0.7.6 → 0.7.7

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## [0.7.7] - 2026-02-20
+
+### Fixed
+- fix(extractor): rewrote importance score calibration in SYSTEM_PROMPT -- per-score definitions (5-10) replace undifferentiated 8-10 band
+- fix(extractor): added signal-cost framing -- 8+ fires real-time cross-session alerts; prompt now uses this as conservative filter
+- fix(extractor): made score 7 the explicit default workhorse; 8+ now requires cross-session justification
+- fix(extractor): added dev-session-observations rule -- verified/tested/confirmed patterns cap at 6 unless result is surprising or breaking
+- fix(extractor): resolved conflict between dev-session cap and explicit memory request rule ("remember this" overrides cap)
+- fix(extractor): removed "verified again today" from score-8 pnpm example to avoid contradicting dev-session rule
+- fix(extractor): added NOT-8 negative examples alongside existing NOT-9 callouts
+- fix(extractor): added 3 non-developer few-shot examples (health at 8, personal at 7, preference at 6) to prevent domain bias
+- fix(extractor): lowered 8+ calibration cap from 30% to 20%
+
 ## [0.7.6] - 2026-02-20
 
 ### Fixed
@@ -22,6 +35,7 @@
 - fix(extractor): increase MAX_PREFETCH_RESULTS from 3 to 5 and lower PREFETCH_SIMILARITY_THRESHOLD from 0.78 to 0.72
 - fix(extractor): increase PREFETCH_CANDIDATE_LIMIT from 10 to 15 for broader elaborative encoding candidates
 - fix(extractor): tighten extractor prompt to suppress near-variant entries already captured in DB
+- fix(extractor): recalibrate importance scoring anchors so routine verifications and test-pass observations default to 6-7; reserve 8+ for cross-session alert-worthy updates
 
 ### Added
 - feat(plugin): signalCooldownMs config - minimum ms between signal batches per session (default: 30000)

package/dist/cli-main.js

@@ -8964,17 +8964,69 @@ If uncertain whether durable, skip.
 
 ## Importance (1-10)
 
-Emit only importance >= 5. Start every candidate at 5; raise only with clear justification.
+Emit only importance >= 5. Start every candidate at 7; lower or raise only with clear justification.
 
-8-10: biographical facts, durable strategic decisions, foundational architecture
-6-7: meaningful project facts, preferences, events that matter beyond this week
-5: borderline but still durable for days/weeks and actionable in future context
-1-4: noise \u2014 do not emit
+Importance scores map to real behavior in the memory system:
+- 8 or higher fires a real-time cross-session signal (an alert to other active AI sessions)
+- 7 is stored silently; no alert fires
+- Below 7 is stored but deprioritized in recall
+
+Use that signal cost as your conservative filter. Ask: "Does someone in another session need to know this RIGHT NOW?" If no, stay at 7 or below.
+
+Score anchors:
+
+10: Once-per-project facts. Core identity, permanent constraints, "never forget this."
+    Example: "This project must never use GPL-licensed dependencies."
+    Example: "The production database password rotation requires manual approval."
+
+9: Critical breaking changes or decisions with immediate cross-session impact.
+    Use for: major architecture reversals, breaking API changes, critical blockers discovered.
+    Example: "agenr embed API changed: model param is now required; all callers must update."
+    Example: "Decided to abandon SQLite-vec in favor of Postgres pgvector - all storage code changes."
+    NOT 9: "we verified signals work" (that is a 6)
+    NOT 9: "tests are passing" (that is a 5-6)
+    NOT 9: "deployed feature X" (that is a 7 event at most)
+
+8: Things an active parallel session would act on if notified right now.
+    Use for: new user preferences discovered, important architectural facts just learned,
+    active blocking issues, key decisions made today that others need to know.
+    Example: "User prefers pnpm over npm for all projects in this workspace."
+    Example: "The chunker silently drops chunks over 8k tokens - callers must split first."
+    If in doubt between 7 and 8, use 7.
+    NOT 8: "we decided to use TypeScript strict mode" (that is a 7 decision)
+    NOT 8: "the user's daughter is named Emma" (that is a 7 biographical fact)
+    NOT 8: "confirmed the import worked" (that is a 6 verification)
+
+7: Default for solid durable facts. Stored, retrievable, no alert.
+    Use for: project facts, preferences (non-critical), completed milestones, stable architecture notes.
+    Example: "agenr stores entries in SQLite with sqlite-vec for vector search."
+    Example: "Completed brain audit. Found 73% noise rate in knowledge base."
+    This is the right score for most extracted entries.
+
+6: Routine durable observations. Worth storing but minor.
+    Use for: dev session observations, test results, routine verifications, minor notes.
+    Example: "Verified that signal emission works end to end in local testing."
+    Example: "Confirmed the import path change did not break CLI startup."
+    Example: "agenr extraction runs in about 2s per chunk on the test dataset."
+
+5: Borderline. Only emit if clearly durable beyond today and actionable in a future session.
+    Example: "Port 4242 is the default for the local test server."
 
 Calibration:
 - Typical chunk: 0-3 entries. Most chunks: 0.
-- 8+ entries: usually 0-1 per chunk
-- If >30% of emitted entries are 8+, you are inflating
+- Score 9 or 10: very rare, at most 1 per significant session, often 0
+- Score 8: at most 1-2 per session; ask the cross-session-alert question before assigning
+- Score 7: this is your workhorse; most emitted entries should be 7
+- Score 6: routine dev observations that are still worth storing
+- If more than 20% of your emitted entries are 8 or higher, you are inflating
+
+Dev session observations rule: Anything in the form "we tested X and it worked", "verified X",
+"confirmed X runs", "X is passing" belongs at 6 unless the result was surprising or breaks
+something. Surprising means a critical bug was found or a breaking change was discovered --
+minor timing differences or slightly unexpected output do not count. Test passes and routine
+verifications are not cross-session alerts.
+Exception: if the user explicitly said "remember this" or "remember that", the explicit memory
+request rule takes precedence and the 6-cap does not apply.
 
 ## Subject (critical)
 
@@ -9038,7 +9090,8 @@ Before emitting EACH entry, all five must be true:
 2. Durable beyond the immediate step
 3. Non-duplicate of another entry in this batch
 4. Importance >= 5 with a concrete reason
-5. Explicit user "remember this/that" requests justify importance >= 7 regardless of content type
+5. Explicit user "remember this/that" requests justify importance >= 7 regardless of content
+   type, including dev session verifications (this overrides the dev-session-observations rule).
 If any check fails, do not emit.
 
 ## Few-Shot Examples
@@ -9122,6 +9175,50 @@ EVENT:
   "source_context": "Brain audit completed and findings documented"
 }
 
+EVENT:
+{
+  "type": "event",
+  "subject": "agenr signal emission",
+  "content": "Verified end-to-end signal emission works in local testing. Entry stored, signal fired, received in OpenClaw session.",
+  "importance": 6,
+  "expiry": "temporary",
+  "tags": ["agenr", "signals", "testing"],
+  "source_context": "Dev session verification of signal feature"
+}
+
+FACT:
+{
+  "type": "fact",
+  "subject": "user penicillin allergy",
+  "content": "User is allergic to penicillin. Must never suggest it or related antibiotics.",
+  "importance": 8,
+  "expiry": "permanent",
+  "tags": ["health", "allergy", "critical"],
+  "source_context": "User mentioned allergy when discussing a doctor visit -- scored 8 not 9/10 because a parallel session needs to know this now (e.g. to avoid recommending antibiotics), but it is not a breaking change or architectural decision"
+}
+
+FACT:
+{
+  "type": "fact",
+  "subject": "user daughter name",
+  "content": "User's daughter is named Emma. She is 8 years old.",
+  "importance": 7,
+  "expiry": "permanent",
+  "tags": ["family", "personal"],
+  "source_context": "User mentioned daughter while discussing weekend plans"
+}
+
+PREFERENCE:
+{
+  "type": "preference",
+  "subject": "meeting time preference",
+  "content": "User prefers morning meetings before noon and avoids late afternoon calls.",
+  "importance": 6,
+  "expiry": "long-term",
+  "tags": ["schedule", "preference"],
+  "source_context": "User mentioned scheduling preference during calendar discussion -- scored 6 not 8 because no parallel session needs to act on this immediately; it is a low-urgency convenience preference"
+}
+
 ### BORDERLINE \u2014 skip these
 
 SKIP: "The assistant read the config file and found the port was 3000."

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agenr",
-  "version": "0.7.6",
+  "version": "0.7.7",
   "openclaw": {
     "extensions": [
       "dist/openclaw-plugin/index.js"
@@ -11,6 +11,13 @@
   "bin": {
     "agenr": "dist/cli.js"
   },
+  "scripts": {
+    "build": "tsup src/cli.ts src/cli-main.ts src/openclaw-plugin/index.ts --format esm --dts",
+    "dev": "tsup src/cli.ts src/cli-main.ts --format esm --watch",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "typecheck": "tsc --noEmit"
+  },
   "dependencies": {
     "@clack/prompts": "^1.0.1",
     "@libsql/client": "^0.17.0",
@@ -54,11 +61,9 @@
     "README.md"
   ],
   "author": "agenr-ai",
-  "scripts": {
-    "build": "tsup src/cli.ts src/cli-main.ts src/openclaw-plugin/index.ts --format esm --dts",
-    "dev": "tsup src/cli.ts src/cli-main.ts --format esm --watch",
-    "test": "vitest run",
-    "test:watch": "vitest",
-    "typecheck": "tsc --noEmit"
+  "pnpm": {
+    "overrides": {
+      "fast-xml-parser": "^5.3.6"
+    }
   }
-}
+}