memory

name	memory
description	Search, pin, decay, and inspect memories. V3 transcript ch.14 — the memory palace skill.
triggers	/memory, /pin, /forget, search memory, what do you know about

Memory

V3 transcript ch.14 — searchable memory palace with importance/salience/recency.

When to use

• /memory <query> — search across memories table
• /pin <id|content> — mark memory as pinned (never decays, injected every session)
• /forget <id> — soft-delete a memory
• "what do you know about X" — natural language search
• /insights — see skills/insights/SKILL.md

Inputs

• query: string — keyword or natural language
• limit: int (default 10)
• agent_scope: optional — filter by agent

Behavior

Search:

1. FTS5 keyword pass on memories.content
2. Embedding similarity pass (Gemini 768-dim, cosine)
3. Merge + dedupe + rank by salience = importance * 0.6 + recency * 0.3 + match_score * 0.1
4. Return top N

Pin:

1. Insert into pinned table with scope
2. Force importance = 1.0 on source memory; bypass weekly decay

Decay (cron, weekly):

1. UPDATE memories SET importance = importance * 0.95 WHERE id NOT IN (SELECT memory_id FROM pinned)
2. DELETE FROM memories WHERE importance < 0.3 AND id NOT IN pinned
3. Audit each change to importance_audit

Output format (search)

MEMORY — "{query}" — {n} hits
━━━━━━━━━━━━━━━━━━━━━━━
[{id}] ({importance}) {kind}: {content}
       agent={agent} · {created_at_human}
       {📌 pinned if applicable}

Tools used

• SQLite (FTS5, regular reads)
• Gemini embeddings 768-dim

Cost notes

• Embedding cost only on memory write (~$0.0001 per memory)
• Search itself is free (local SQLite)

See also

• memory/inject.ts — how pinned + top-importance memories are loaded at session start
• memory/extractor.py — Gemini washing machine that creates memories from transcripts
• memory/consolidator.py — merges near-duplicates