Stop explaining the same thing twice

Memory gives a future session useful learned context without pretending history is the same as current truth. Save the smallest durable fact when it is learned, then verify mutable state before acting on it.

Save knowledge, not live state

Use Memory for a decision, preference, recurring lesson, project motivation, or safe reference that another session would otherwise have to rediscover. Keep standing rules in AGENTS.md, current architecture in source, and unfinished intent in a Plan.

Direct proof: Start a fresh session with a related ordinary request. It should recall the named decision with provenance, while still checking live source before a consequential change.

What changes in future sessions

1. Faster starts. Relevant prior decisions arrive before repeated discovery.
2. Fewer regressions. Hard-won lessons remain available to later sessions.
3. Better judgment. Provenance and scope help the agent decide when a memory applies.

Mission: preserve one useful lesson

1. Identify a fact that will matter beyond the current session.
2. Check whether Memory already records it accurately.
3. Save one concise record with the reason, scope, and when it applies.
4. Start a later related session and inspect the recalled source.
5. Verify any mutable fact against the current authoritative system.

Keep the source boundaries clear

Memory is not a secret store, task tracker, source-code index, or substitute for live verification. Do not save passwords, tokens, one-time codes, transient status, or facts that the repository answers directly. Supersede wrong memories instead of leaving two conflicting versions.

Related references

• The working agreement
• Work plans
• How to work with Axis

Put this to work: Set the project agreement that Memory should support—not replace.