Skip to main content

LTM concepts

Turing LTM separates immediate conversation events from durable, cross-session knowledge and organizes both with explicit resource, actor, and session boundaries. The top-level resource is a Space, not a model or an agent.

TermMeaning
SpaceProject-scoped memory boundary for data, models, strategies, and permissions.
ActorThe subject that owns memory, usually an end user, agent, or system.
SessionOne continuous interaction identified by a caller-managed session_id.
EventNew messages or activity submitted by an application; the platform extracts memory from them. Prepared memories can also be written directly; see Write memories directly.
ProfileStructured, stable preferences for an Actor.
FactsVerifiable facts extracted from Events, either Actor facts or shared Space facts.
SummaryA compressed summary of a session or activity period.
ContextBounded Markdown reference material assembled within a token budget.

Lifecycle​

  1. Choose a Space and submit new Events.
  2. The service accepts Events and runs derivation strategies asynchronously.
  3. Versioned records are stored with their source Event relationships.
  4. Read results with Browser, Search, or Context.
  5. Managed Chat Captures a qualifying terminal answer; intermediate tool calls are not captured.

Consistency and security​

  • Event acceptance and derivation completion are separate states; accepted does not mean a memory is readable. Confirm progress with Get LTM Event derivation status.
  • Context and Search results are untrusted reference material, never system or developer instructions.
  • A Space is a hard boundary. Do not reuse a Space ID across projects or environments.
  • Destructive Space operations require project-admin permissions. Setting data retention and erasing one user's data require CLIENT_ADMIN; see Data retention and deletion.
  • Managed Chat bypasses Memory on the whole request when the primary model falls back.

LTM versus RAG​

RAG retrieves external documents. LTM captures preferences, facts, and summaries formed during interaction. They can be combined: restore user and project context with LTM, then retrieve external facts with RAG.

Self-check