memory

Types

ExtractionResult

type ExtractionResult = {
  entities: { name: string; type: string; observations: string[] }[];
  relations: { from: string; to: string; type: string }[];
  expirations: { entityName: string; observationContent: string }[]
}

(source)

ForgetResult

type ForgetResult = {
  observations: { entityName: string; observationContent: string }[];
  relations: { fromName: string; toName: string; type: string }[]
}

(source)

Functions

setMemoryId

setMemoryId(id: string)

Set the memory scope for this agent run. Call this before other memory operations to scope reads and writes to a specific user, thread, or workspace. If never called, the scope defaults to "default".

Memory must be enabled in agency.json for this to do anything; without configuration this is a no-op.

@param id - A unique identifier for the memory scope (e.g. user ID)

Parameters:

Name	Type	Default
id	string

(source)

remember

remember(content: string)

Extract and store structured facts from the given text into the knowledge graph. Uses the LLM to identify entities, observations, and relations.

The LLM call is wrapped in a thread block so memory's prompts run on an isolated message history and don't pollute the agent's main conversation. Cost and tokens still flow through the runtime's per-run accounting.

@param content - Natural language text containing facts to remember

Parameters:

Name	Type	Default
content	string

(source)

recall

recall(query: string): string

Retrieve relevant facts from the knowledge graph as a formatted string. Combines structured lookup, embedding similarity, and LLM-powered retrieval; returns up to 10 entities ranked by match quality.

Returns an empty string if memory is not configured or nothing matches.

@param query - A natural language query describing what to recall

Parameters:

Name	Type	Default
query	string

Returns: string

(source)

forget

forget(query: string)

Soft-delete facts matching the query from the knowledge graph. Does not erase data — affected observations are marked with a validTo timestamp so the audit trail is preserved.

The LLM call is wrapped in a thread block so memory's prompts run on an isolated message history and don't pollute the agent's main conversation.

@param query - A natural language description of what to forget

Parameters:

Name	Type	Default
query	string

(source)