memory

To keep context across sessions, persistent cross-session memory stores past decisions and patterns in a local SQLite database, injecting prior cases into workflows at runtime to influence current tasks.

Local vector embedding memory works by using sqlite-vec to index 384-d vectors from fastembed, allowing an agent to query and retrieve similar past cases or corrections from an on-disk SQLite DB.

Yes, you can run cross-session memory retrieval completely offline because the system is local-first and deterministic, relying on an on-disk SQLite database and offline fastembed vector embeddings.

The best way to prevent memory poisoning is to apply deterministic write and update rules with quarantine and promotion stages, ensuring only validated patterns are saved to the local database.

You can seed and query past cases using a provided CLI that interacts with the on-disk SQLite database, allowing you to manually input prior decisions and retrieve them via vector similarity.

You need persistent memory for brainstorm and audit phases when you want to recall prior decisions, patterns, and outcomes to make context-aware choices and streamline ongoing work.