How it works

Three things to know

Memory Layer has many features, but its mental model is three ideas. If you understand this page, you can use the product; everything else is optional depth.

1. Memories — durable facts, not chat logs

When work completes, memory remember captures it. Curation distills the capture into short, citable memories: a durable claim, its scope, and the evidence behind it (files, commits, notes). Raw conversation is never stored as memory — only what stays true tomorrow.

memory remember --project app --title "Switched auth to sessions" \
  --summary "JWTs replaced with server-side sessions" \
  --note "Session cookies are HttpOnly; TTL 24h; see auth/session.rs"

2. Retrieval — ask, get a cited answer or an honest refusal

memory query searches those memories three ways at once — keywords, meaning (embeddings), and the code graph — and synthesizes a short answer that cites the exact memories it used. When the evidence is weak, it says so (insufficient_evidence) instead of guessing. memory resume uses the same retrieval to build a briefing that gets you back into flow after time away.

memory query --project app --question "How does auth work now?"
memory resume --project app

3. Reinforcement — memory maintains itself

Every retrieval strengthens the memories that earned it; activation spreads to linked memories and decays over time, exactly like the ACT-R cognitive model. Frequently used knowledge stays sharp; hot memories get validated against your actual code; stale claims lose rank and can be superseded — with every durable change gated behind human review.

You don't run reinforcement. It happens because you use the system.

That's the loop

remember → query/resume → (automatic) reinforcement. Three commands cover daily use — try them live with memory tour.

When you want more depth

© 2026 Olivier Van Acker (3vilM33pl3). Memory Layer is AGPL-3.0-or-later with commercial licensing available.

On this page