Three meanings of "memory"
- In-conversation — message history within one session. The model sees it on every turn until it's compacted out.
- Cross-conversation — persistent facts the user wants the model to recall across sessions. Implemented in code, stored in a database or vault.
- Native memory — provider-managed memory features (OpenAI memory tool, Anthropic memory beta, ChatGPT-style implicit recall).
What to remember
- User preferences that persist (timezone, language, formatting choices).
- Stable facts about the user's project or domain.
- Decisions made in earlier sessions that shape later ones.
What not to remember
- Ephemeral task state (current step in a workflow). Use scoped session state.
- Sensitive data the user wouldn't want recalled later (payment, health).
- Conjectures and one-off speculations.
Who decides
Default to: the user explicitly opts in, or your system explicitly classifies a fact as durable. Implicit memory of everything is a privacy bug, not a feature.