syiibfs-hash/cc-agent-brain/cc-agent-brain-memory-distill

CC Agent Brain Memory Distillation

Use when:

• a session contains stable user preferences
• an active project changes state
• repeated patterns or corrections appear
• the Stop hook suggests distillation, or candidates shows a backlog

Step 1 — inspect what accumulated

python .claude/hooks/cc-agent-brain/brain_store.py candidates --days 3
python .claude/hooks/cc-agent-brain/brain_store.py search "<topic>" --k 10

Step 2 — write versioned memories

python .claude/hooks/cc-agent-brain/brain_store.py memory add \
  --title "..." --body "..." \
  --layer semantic|procedural|working|episodic \
  --confidence high|medium|low \
  --tags "tag1,tag2" --source "wiki/..."

Layers:

• semantic — stable facts, preferences, project state (the default)
• procedural — reusable workflows, checklists, prompt patterns
• working — short-lived active-task context
• episodic — dated events that must keep their time anchor

Step 3 — supersede, never overwrite

When a new fact replaces an old one, write the new memory with --supersedes <old-id>. The old memory is kept as superseded with valid_to set, and leaves active retrieval — this preserves an auditable history instead of silently rewriting.

When a fact is simply no longer true, expire it:

python .claude/hooks/cc-agent-brain/brain_store.py memory expire <id>

Step 4 — refresh projections

python .claude/hooks/cc-agent-brain/brain_store.py sync-md

Also keep memory/MEMORY.md and the relevant memory markdown files aligned.

Do not distill

• raw conversation, temporary tasks, uncertain guesses
• secrets or credentials (the engine redacts common key shapes at capture time,

but do not deliberately store sensitive material)

• information likely to expire soon — use working layer instead

Memory item contract

Every memory carries: id, layer, title, body, tags, confidence, status, valid_from, valid_to, source, supersedes, timestamps.