Metadata filter (v0.2.4)

Pass metadata_filter=<urlencoded JSON object> on GET /api/memories or in the bulk-search body. Translates to Postgres' JSONB @> (containment) operator.

# All memories where metadata = { speaker: "sarah" }
curl -H "Authorization: Bearer mnu_..." \
  "https://mnueron.com/api/memories?metadata_filter=$(printf %s '{"speaker":"sarah"}' | jq -r @uri)"

# SDK
mem.search("our pricing approach", metadata_filter={"speaker": "sarah"}, k=5)

Notes

• Top-level keys only in v0.2.4. Nested matching works via JSON equality ({ "user": { "id": "abc" } } matches the whole nested object).
• All filter conditions are AND'd together — every key=value pair must match for the memory to be returned.
• Stacks with q, namespace, date filters in any combination.

Schema patterns that make this fast

Use predictable metadata keys at save time so filters don't return surprises:

mem.save(
    content=decision.text,
    namespace=f"meeting-{meeting_id}",
    tags=["decision"],
    metadata={
        "speaker": decision.speaker,   # short string
        "ts_s": decision.timestamp,    # number
        "category": "pricing",         # enum-like string
    },
)

Then queries like metadata_filter={"category": "pricing"} are trivial.