Persistent Memory

Persistent memory carries useful context beyond one conversation. It can reduce repeated discovery, but it also creates a new failure mode: stale or low-provenance context can silently shape future work.

Distinguish Memory from Instructions

Project instructions are human-authored repository contracts. Memory is accumulated recall: preferences, discoveries, recent context, or agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose-authored notes. Required team rules belong in version-controlled instructions or documentation, even when a product can remember them.

Claude Code documents repository-scoped local auto memory alongside CLAUDE.md. Local Codex clients use a separate local memory store and controls, while hosted ChatGPT behavior is not the same store. Copilot CLI documents memory tooling and enablement, but the reviewed source does not establish one shared memory contract across every Copilot surface.

A Governed Memory Layer

Every durable memory should answer:

• Where did this claim come from?
• When was it observed or reviewed?
• What scope does it apply to?
• Who or what may update it?
• How is contradiction or staleness detected?
• Can a user inspect, correct, export, or delete it?

Use stable identifiers and source links for facts that drive decisions. Separate observed state from inferred preference. Expire project-status memories aggressively; retain durable architectural decisions only while their evidence remains valid.

Operational Rule

Treat memory as a recall accelerator, not a source of authority. Revalidate load-bearing facts against code, configuration, or a current primary source before acting. Never rely on model memory alone for release authorization, credentials, customer commitments, or production state.

The deeper lenses retain Claude Code auto-memory mechanics. The governance model transfers; storage paths, loading thresholds, and controls do not.