Memory

What an agent remembers and who else can read it — one org-scoped store for the team wiki and agent memory, plus a per-user recall surface.

After this page you know where to put something an agent should remember, and who will be able to read it back.

There is no capability called Brain. There are two real memory surfaces with different scopes, and picking the wrong one is the usual way a memory ends up invisible to the agent that needed it.

Org memory: knowledge

knowledge is the one that matters for teams. Wiki pages a person writes, memories an agent files, and documents a connector ingests are one document store, indexed into that org's own vector namespace on every save.

That is the whole design decision: human wiki and agent memory are one store indexed once, so an agent retrieves exactly what the team can read — no more, and no less.

curl -X POST https://api.hanzo.ai/v1/knowledge/search \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query":"how do we handle refunds"}'

This is the retrieval entry point: an agent asking what does this org know about X, answered from the org's own namespace. Both the collection and the payload filter are pinned to the caller's org, so a query cannot reach another tenant's documents even by accident. If the index is unreachable it answers empty with degraded: true rather than a 5xx — a degraded answer is a fact the caller can act on; a 500 is not.

Beyond search: /v1/knowledge/import for bulk, and /v1/knowledge/connectors to keep an outside source in sync.

Personal memory: recall

The ai capability carries a per-user memory surface, for context injection into a conversation:

POST /v1/ai/memory/remember
POST /v1/ai/memory/recall
POST /v1/ai/memory/search
GET  /v1/ai/memory/list
GET  /v1/ai/memory/facts

recall ranks semantically when you pass a query and returns the most recent when you do not. Everything here is scoped to the authenticated user, not the org — it is the assistant remembering you, not the company remembering something.

Use it for preferences and conversational continuity. Use knowledge for anything a colleague should also be able to find.

The rest of the retrieval family

Memory is one of five stores, and they compose rather than compete:

	What it holds	Ask it
knowledge	documents + vectors, per org	what do we know about X
index	a full-text index	which document contains this word
search	both of the above, fused	one ranked answer
graph	entities and relations, as assertions	how is X connected to Y
dataset	versioned immutable snapshots	what did this look like then

search is usually the one you want: POST /v1/search fuses the lexical index and the knowledge vector index into a single ranked result, so you do not choose between keyword and meaning.

graph is worth knowing about because of what it stores: not facts, but assertions — somebody, at some moment, from some evidence, asserted that this thing stands in that relation to that other thing. That makes disagreement representable instead of a write conflict.

Knowledge — the org store.
Search — fused retrieval.
Agents — what does the remembering.

Org memory: knowledge

Personal memory: recall

The rest of the retrieval family

Next

On this page