Skip to main content
Searches the agent’s long-term memory — a temporal knowledge graph of facts learned from past tasks — for information relevant to a query. Display name “Search Memory”. Off by default; enable it per agent. This is a parallel memory system to the agent’s learned rules. It stores entities (vendors, documents, people) and the facts/relationships between them, tracks how those facts change over time, and retrieves via hybrid semantic + keyword + graph search. Facts learned from confirmed/approved rules are marked approved and surfaced first.

​
Authentication and enablement

  • No external integration. Backed by the internal graphiti-sidecar service.
  • Dark by default: available only when graph memory is enabled for the deployment (MEMORY_GRAPHITI_ENABLED) and the tool is added to the agent’s tools. Not seeded into new agents.
  • Scoped to the running agent’s group — an agent only ever searches its own group’s memory.

​
Inputs

  • query (required): what to recall — a question, topic, or entity. Max 10,000 characters.
  • max_facts (optional): 1–50, default 10.

​
Output

A ranked list of facts. Each fact includes its text, whether it is approved, its source (e.g. observation or approved_rule), and its validity window (valid_at / invalid_at). Approved facts are ordered first.

​
Limits and side effects

  • Read-only. Does not modify memory; ingestion happens separately (from task activity and approved rules).
  • Returns facts, not documents or files — use retrieve_documents for verbatim document chunks.
  • Results reflect what has been ingested so far; a fresh deployment returns nothing until memory is populated.

​
Expected errors

  • Empty or oversized query.
  • Missing agent-group context.
  • Memory backend unavailable.