Ask & answer synthesis

/ask turns a natural-language question into a grounded answer: the node runs hybrid recall over the profile’s memories, then its built-in assistant synthesizes a prose answer citing the memory ids it rests on.

memoturn ask acme alice "what can this user eat?"
# vegan — switched from vegetarian in 2026
# sources: mem_9f2c…

asked = alice.ask("what can this user eat?")
# {"answer": "...", "sources": ["mem_9f2c…"], "memories": [...]}

answer is null when nothing relevant was recalled — the assistant never invents an answer from an empty result. The recalled memories ride along in memories for attribution or display, and the same filters as recall apply (types, session_id, source, k, include_superseded).

Opting a node in

Answer synthesis is a per-node opt-in and stays off the write path:

env var	role
MEMOTURN_ASSISTANT_API_KEY	enables /ask; falls back to MEMOTURN_EXTRACT_API_KEY if unset
MEMOTURN_ASSISTANT_MODEL	optional model override

A node without a key keeps every other endpoint working; /ask returns 503 with code unconfigured.

The fallback pattern

Because the 503 is machine-readable, clients degrade gracefully — recall, then synthesize with whatever model the agent already has:

try:
    answer = alice.ask(question)["answer"]
except MemoturnError as e:
    if e.code != "unconfigured":
        raise
    memories = alice.recall(query=question)["memories"]
    answer = my_model.summarize(question, memories)  # bring-your-own synthesis

The memory-agent example in the repository runs this exact loop, and the MCP server’s memory_ask tool surfaces the same behavior to agent frameworks.

Governance

ask is AI egress: the question and recalled memories leave the database for the assistant model. Namespaces can deny it wholesale with the ai_egress.ask = deny policy rule, and audit streams record each call’s metadata when enabled — see security and the governance policy docs.