Project memory

C2 can remember stable project facts and earlier outcomes across sessions — even when the next session uses a different agent provider. The memory stays local in C2's SQLite store and is scoped to the exact project path.

Open Settings → Memory to control or inspect it.

What the layers mean

Layer	Meaning
L0 · Raw transcript	Your existing C2 conversation history. It is searched, not duplicated.
L1 · Stable notes	Constraints, preferences, facts, decisions, and notes you add yourself.
L2 · Work episodes	A compact request/outcome record from a completed turn.
L3 · Project profile	A conservative roll-up made after the project has at least three stable notes.

Every derived item keeps a pointer back to the session and transcript part that produced it. A manual note is labeled manual.

Controls

• Enable memory is the master switch.
• Learn from completed turns controls new L1–L3 capture.
• Recall into prompts controls whether a small relevant memory block is sent with new prompts.
• Learn from external context controls whether turns that used tools, files, MCP, browser context, images, referenced chats, or recalled memory may create durable memory.
• Search finds records across all four layers. Chinese text uses bigram matching, so a query does not need spaces between words.
• Pin a stable L1 note to keep it in the bounded automatic-recall pocket.
• Forget an editable memory to deactivate it. The toast's Undo action restores it.

Turning memory off does not erase data. That makes it safe to pause recall without losing the project's history.

Stable L1 facts do not become durable immediately. They wait in a visible candidate count for 30 minutes after the session's latest completed turn, then background maintenance promotes them and refreshes L3. L2 work episodes remain immediate and source-linked.

The composer also controls each session independently:

• Memory on recalls and learns.
• Recall only recalls but does not learn from the session.
• Private session neither recalls nor learns.
• Learn only learns without sending existing memory into the session.

What reaches the model

C2 sends at most 12 L1 notes, 3 earlier L2 episodes, 1 L3 profile, and — only when you explicitly ask about earlier work — 3 raw L0 excerpts. Current-session episodes are excluded.

Recalled material is labeled as untrusted reference data. Your current request and repository rules always take precedence, and recalled transcript text is never treated as an instruction by itself. The injected block is not saved back into the transcript, so it is not recaptured as user-authored input.

Each turn that receives recalled context shows a Memory used receipt. Expand it to inspect the exact items, source pointers, and estimated prompt tokens. The receipt is stored separately from the conversation, so transparency does not contaminate future transcript recall.

Capture ignores expanded file contents, rule bodies, and skill bodies. It also redacts common secret shapes before saving derived text, though you should still avoid placing credentials in a manual memory.

Memory and provider sessions

Project memory complements provider-native session context; it does not replace it. Continuing one Codex or Claude Code session still uses that provider's ACP session state. Project memory supplies a small, inspectable bridge when you start another session or use a different provider.