feat(memory): ICM slices 1-3 — memoirs + feedback + importance/decay

[NagyVikt]

Summary

Unifies the ICM integration on one branch. The three slices from docs/icm-integration-plan.md (memoirs, feedback, importance + temporal decay) were drafted in parallel, all targeting migration 015 — this PR lands them in canonical order on a single, testable branch.

Slice 1 — memoirs (typed knowledge graphs)

• Tables: memoirs, memoir_concepts, memoir_relations; FTS5 mirror over name/content/labels with the standard (ai, ad, au) triggers.
• Nine relation types (part_of, depends_on, related_to, contradicts, refines, alternative_to, caused_by, instance_of, superseded_by); self-loops forbidden; (source, target, relation) unique.
• MemoryStore.{createMemoir, addConcept, refineConcept, linkConcepts, searchConcepts, inspectConcept, listMemoirs} — concept content routes through prepareMemoryText (same redact → compress pipeline as observations).
• Seven MCP tools (memoir_create / list / add_concept / refine / link / search / inspect) with progressive disclosure: memoir_search returns compact hits; memoir_inspect returns body + BFS neighbourhood.

Slice 2 — feedback record/search/stats

• Table: feedback (topic, prediction, correction, optional context, importance, created_by) + feedback_fts virtual table.
• MemoryStore.recordFeedback / searchFeedback / getFeedback / feedbackStats — prediction/correction/context bodies pass through prepareMemoryText, so the compression invariant holds at the write boundary.
• MCP: feedback_record / feedback_search / feedback_stats with compact-hit progressive disclosure.

Slice 3 — observation importance + temporal decay

• New columns on observations: importance (critical | high | medium | low, default medium), access_count, last_accessed_at, weight.
• Storage.recordAccess(ids): single transactional UPDATE that increments access_count, stamps last_accessed_at, and recomputes weight. Critical/high pin to baseWeight(importance); medium/low decay as baseWeight / (1 + access_count * 0.1).
• Storage.pruneLowDecay({ minWeight }) + Storage.countLowDecayCandidates: delete / count medium-or-low rows below the threshold. Critical/high are never affected.
• MCP: search, semantic_search, get_observations carry importance + weight in each row (additive). task_post accepts optional importance.
• CLI: colony memory prune [--min-weight <n>] [--dry-run].

Migration ordering

File	Version	Slice
015-icm-memoirs.ts	15	1
016-icm-feedback.ts	16	2
017-icm-importance-decay.ts	17	3

SCHEMA_SQL keeps every ICM table as CREATE TABLE IF NOT EXISTS (fresh DBs); existing DBs pick up slice 3's column additions via COLUMN_MIGRATIONS. The numbered migration files remain canonical references — the runtime path stays SCHEMA_SQL + COLUMN_MIGRATIONS.

Write-amplification strategy (slice 3)

Debounced 50ms batch via setTimeout: read paths queue returned ids into a Set, the first queue kicks a setTimeout, and the timer flushes the coalesced set in one transactional UPDATE. Every access is counted (unlike threshold sampling), write amplification is capped at one UPDATE per ~50ms hot window, and MemoryStore.close() deterministically flushes pending ids. unref() keeps the timer from holding Node alive. Readonly stores skip the buffer entirely.

Verification

pnpm typecheck                              # green across 17 workspaces
pnpm --filter @colony/mcp-server test       # 302 passed
COLONY_SKIP_PERF=1 pnpm --filter @colony/storage test
# 193 passed | 1 skipped (perf, gated per its own COLONY_SKIP_PERF docstring)
pnpm --filter @colony/core test             # 259 passed; one pre-existing
# `savings-reference.test.ts` failure also fails on the 2d831bb baseline before
# this merge — unrelated to ICM.

Includes the slice 3 perf-smoke receipt that was already in the original PR body:

[importance-perf] n=100 mean=21.77ms p95=41.99ms

Reference

• docs/icm-integration-plan.md
• Upstream: https://github.com/rtk-ai/icm.git

🤖 Generated with Claude Code