lesson-recall default minConfidence (0.1) + decay + auto-prune make most auto-consolidated lessons invisible, with no config override

[nicofirst1]

On v0.9.27 our store had 439 lessons, but default memory_lesson_recall calls returned only a handful, which reads as "agentmemory has almost no lessons." Actual distribution: 438 of 439 active lessons sit at confidence 0.05-0.1, below the default recall threshold - only 1 lesson (0.2%) is visible to a default-parameter recall.

Mechanism (all verified in source):

• mem::lesson-recall defaults minConfidence to 0.1 (src/functions/lessons.ts:101), while a second recall path defaults it to 0 (lessons.ts:157-161) - the two defaults are inconsistent.
• Lessons decay at decayRate: 0.05, so auto-consolidated lessons drift below 0.1 and vanish from default recalls.
• lessons.ts:242: if (lesson.confidence <= 0.1 && lesson.reinforcements === 0) deletes the lesson - so a lesson sitting at the visibility floor is also on the verge of silent deletion.
• src/config.ts has no AGENTMEMORY_* variable for lesson confidence, decay, or minConfidence (only CONSOLIDATION_DECAY_DAYS, which is unrelated).

Ask:

1. Align the two recall defaults (either both 0 or both 0.1).
2. Add a config knob (e.g. AGENTMEMORY_LESSON_MIN_CONFIDENCE) or lower the recall-tool default.
3. Surface "N lessons hidden below minConfidence" in the recall response so users know the store is not empty.

Workaround we use: pass minConfidence: 0 explicitly on every recall.

Related but distinct: #274 (lessons not created in consolidation paths), #457/#458 (context injection), #831 (qualityScore heuristics).

[nicofirst1]

Following up on this on v0.9.28, where the same store has now grown to 1354 lessons. The symptom is worse at scale, and digging into it I think the threshold/decay framing here is treating a symptom - the deeper cause is that lesson confidence has no counterforce to decay at all. Decay is one-directional.

Data (v0.9.28, 1354 lessons)

• 1349 / 1354 sit below confidence 0.1 (invisible to a default recall).
• 0 / 1354 have ever been reinforced past their creation minute - every lastReinforcedAt equals createdAt.
• reinforcement-count distribution: 916 lessons at 0, 320 at 1, the rest scattered higher - but those counts all share the creation timestamp, i.e. they came from within-batch dedup, never from later use.

Mechanism (verified in the v0.9.28 bundle)

The thing that raises confidence is reinforceLesson():

lesson.confidence = Math.min(1, lesson.confidence + 0.1 * (1 - lesson.confidence));
lesson.lastReinforcedAt = now;

It has exactly two callers, and neither is in the recall/use path:

1. mem::lesson-recall is read-only. It filters by confidence >= minConfidence, scores, writes a single lesson_recall audit event, and returns. No kv.set on the lessons, no reinforceLesson. So retrieving and using a lesson never raises its confidence - the usage signal is captured in the audit log and then discarded.
2. mem::lesson-save calls reinforceLesson only when re-saving byte-identical content - but it keys lessons as fingerprintId("lsn", ...) while consolidation keys them as fingerprintId("lesson", ...). Different prefix = different id, so a manual save can never match the auto-consolidated lessons to reinforce them.
3. mem::lesson-strengthen also calls reinforceLesson, but it requires an explicit lessonId and is not exposed as an MCP tool (the registry surfaces memory_lesson_recall and memory_lesson_save, no _strengthen). Nothing auto-invokes it after a recall.

Meanwhile the consolidation merge path does touch lessons on re-import, but it only does reinforcements: (existing.reinforcements||0)+1 and stamps lastReinforcedAt, spreading ...existing - it never re-raises confidence. So the counter you see is a dedup tally, not a usage/strength signal.

Net: auto-lessons are born at confidence 0.4, decay 0.05/week toward the 0.05 floor, cross under the 0.1 recall threshold in ~7 weeks, and nothing in normal operation ever pushes them back up. Aligning the two minConfidence defaults or adding an AGENTMEMORY_LESSON_MIN_CONFIDENCE knob (both worthwhile) would make decayed lessons visible, but they'd still be decaying to the floor with no reinforcement - a genuinely useful lesson recalled 50 times looks identical to one never used again.

Suggested additional fix

Reinforce on recall: have mem::lesson-recall call reinforceLesson() (or a gentler variant) on the lessons it actually returns, so retrieval-and-use becomes the reinforcement signal decay is meant to be balanced against. Bonus cleanups: reconcile the "lsn" vs "lesson" fingerprint prefixes so manual save can dedup against consolidation lessons, and either expose lesson-strengthen as an MCP tool or wire it into the recall path.

Happy to open this as a separate issue if you'd prefer to keep #1065 scoped to the visibility half.