Remembered context

Control the information carried between conversations and inspect memory processing.

Memory helps Mellow recover useful context without putting every past conversation into every request. Use it for preferences and continuity. Use Knowledge for source documents, and an agent database for records that need explicit fields and queries.

Start with the processing model

Memory extraction uses the configured Core Model. A ready chat model does not necessarily mean the Core Model is ready. Open the model settings, check the background model, then inspect Memory for pending work or errors.

A local Core Model keeps the extraction request on the Mac. A remote Core Model sends the material needed for extraction to its provider. Likewise, retained memory may later become context for a remote chat model. Local storage and local inference are separate choices.

What is retained

Identity overrides are treated differently from retrieved context: they are small, explicit instructions intended to remain available. Keep them concise. “Use British spelling in my writing” is clearer than a long mixture of current tasks and personal history.

Recall is selective

Mellow evaluates whether a message needs remembered context. A routine question may need none. A question about a previous decision may retrieve a summary or fact. Asking for exact wording requires the original material rather than relying on a compressed recollection.

The default relevance gate is heuristic. Configuration also supports a classifier-assisted gate and a gate-off mode for always attempting recall. Gate-off is not the same as disabling Memory; it changes the decision about retrieval.

Default processing controls

These are implementation defaults, not a promise that extraction completes at an exact wall-clock time. Availability, pending work, and model execution affect completion.

Per-agent and project scope

An agent has its own memory scope. Project chats additionally use project-scoped context across participating agents. Disabling an agent's personal memory does not mean a conversation inside a project is free of project context. The overall Memory switch governs the system.

When diagnosing an unexpected recollection, check the active agent, project membership, identity overrides, and attached knowledge. Those are distinct sources of context.

Correct, review, or forget

Open Memory and select the relevant scope. Review summaries before treating them as permanent facts. Use overrides for an explicit preference; do not continually repeat a correction in unrelated chats and assume every previous record has been replaced.

Use Sync to process pending updates and Run Now for consolidation where those controls are available. Cleanup can merge or remove stale material; it is not a substitute for explicitly deleting data you no longer want retained. Review the scope of a destructive action before confirming it.

Troubleshooting recall

The extraction pipeline limits repeated parse failures and bounds very long inputs. Short, explicit decisions are easier to retain accurately than a large undifferentiated paste. For algorithm and storage details, see Memory internals.