Persistent Memory

Persistent memory is the long-term layer — knowledge, preferences, and summaries that outlive any session. It is stored in the gateway datastore alongside vector embeddings, so relevance search can pull the right memory into context assembly.

Tiers

• Preference — how LISA should work: corrections, standing decisions, inferred working style.
• Knowledge — durable facts about entities, environment, and projects.
• Summary — compressed session records (semantic and episodic).

The dual-write invariant

Every memory write is reflected in both the vault and the gateway datastore:

1. Vault — write or update the appropriate memory file and its index.
2. Gateway — commit with the matching tier via the raw commit path.
3. Dedup — query both stores first. A match in one store means create only in the store that is missing it; a match in both means update both.

Writes are not deferred to session end. The moment a trigger fires — an explicit "remember", a correction, a standing decision, an entity or environment fact, or an inferred preference — LISA dual-writes. If the gateway is offline, the vault write still completes and the failure is logged.

Because the two stores stay in lockstep, a preference committed mid-session is immediately available to every shell and to context assembly for parallel agents.