Mental Model

Persistent memory is a recall layer that influences future model context. It is neither the current conversation nor the authoritative repository contract.

Five Dimensions

1. Author — human, 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, or derived system process.
2. Scope — user, repository, role, workspace, or hosted account.
3. Store — local machine, checked-in file, or managed service.
4. Retrieval — always loaded, indexed, or recalled on demand.
5. Lifecycle — reviewed, corrected, expired, exported, or deleted.

Claude Code documents local repository-scoped auto memory alongside CLAUDE.md. Local Codex clients use their own local memory store, distinct from hosted ChatGPT memory behavior. Copilot CLI documents memory controls, while one cross-surface Copilot memory contract is not established by the reviewed CLI source.

Recall Versus Authority

Memory may say, “This repository usually uses command X.” The repository and current build configuration decide whether X is still correct. Keep non-negotiable team rules in version-controlled instructions and verify volatile facts against their source before action.

The crux is provenance: a memory without origin, scope, and review date is an unpriced source of future error.