mirror of
https://github.com/hanzoai/agents.git
synced 2026-10-30 18:48:36 +00:00
zeekayandHanzo Dev
b78003dbc3
BaseMemoryBackend stores one record per scope and key in one collection, over
Base's /v1/collections/{name}/records. MEMORY.md and
agent_memory.collection.json create the collection.
Records use the scope names the control plane and the other SDKs use, so Go's
user scope is filed as actor. A create that loses a race to the unique index
updates the record that won: the last write wins.
The in-memory backend's similarity search returned nothing. It now ranks by
cosine and honours filters, through the same math as the Base store.
Co-authored-by: Hanzo Dev <[email protected]>
1.5 KiB
1.5 KiB
Agent memory in Hanzo Base
BaseMemoryBackend keeps an agent's memory in Base, so its state lives in the
same single binary as everything else it runs beside. One record per scope and
key, in one collection.
Create the collection
Import the schema once, as a superuser:
curl -X PUT http://127.0.0.1:8090/v1/collections/import \
-H "Authorization: $(hanzo auth token)" \
-H 'Content-Type: application/json' \
--data @agent_memory.collection.json
Base's admin UI imports the same file from Settings.
Nothing creates it on first write.
Two things in that schema are load-bearing:
- The unique index over
(scope, scope_id, mkey)is what makes one key one record. Without it a losing race writes a second record at the same key and reads start depending on which one Base returns first. - The rules are null, so out of the box the collection is reachable by a superuser alone. Open them deliberately, per deployment.
Use it
memory := agent.NewMemory(
agent.NewBaseMemoryBackend("http://127.0.0.1:8090", token, "agent_memory"),
)
memory.Set(ctx, "tone", "plain")
NewMemory(nil) gives the in-memory backend instead, which is right for a test
and loses everything when the process ends.
Similarity search
The caller computes embeddings. SearchVector ranks the scope's vectors by
cosine in the client, which suits an agent's working set, not a corpus. Vectors
of another width came from another model and are skipped.