agent-memory

Agent Memory

Governed operational memory for agent runtimes, with provenance, review, recall traces, and audit trails.

flowchart LR
  Runtime["Agent runtime<br/>OpenClaw, Codex, local agents"] --> Recall["Recall request"]
  Recall --> OB1["OB1 Agent Memory"]
  OB1 --> Memories["Scoped memories<br/>provenance + use policy"]
  Memories --> Runtime
  Runtime --> Writeback["Compact write-back"]
  Writeback --> Review["Human review"]
  Review --> Future["Future recall"]
  Future --> Runtime

Loading

What It Does

This schema adds sidecar tables that let Open Brain store agent-created operational memory safely. The core thoughts table remains the content store; agent memory records add provenance, confidence, scope, use policy, review status, source references, recall traces, and audit events.

Prerequisites

• Working Open Brain setup (guide)
• Supabase project with the core thoughts table
• The core dedupe setup from Step 2.6 is recommended

Credential Tracker

AGENT MEMORY -- CREDENTIAL TRACKER
--------------------------------------

SUPABASE (from your Open Brain setup)
  Project URL:           ____________
  Secret key:            ____________

--------------------------------------

Steps

Open Supabase SQL Editor, paste the contents of schema.sql, and run it.

Done when: Table Editor shows agent_memories, agent_memory_recall_traces, agent_memory_recall_items, and agent_memory_audit_events.

Run this query:

SELECT column_name, column_default
FROM information_schema.columns
WHERE table_name = 'agent_memories'
  AND column_name IN (
    'can_use_as_instruction',
    'can_use_as_evidence',
    'requires_user_confirmation',
    'review_status'
  );

Done when: instruction defaults to false, evidence defaults to true, confirmation defaults to true, and review defaults to pending.

Deploy the runtime API from ../../integrations/agent-memory-api/.

Done when: GET /health on the deployed API returns {"ok":true}.

Expected Outcome

After applying this schema, OB1 can store agent memories as governed records instead of raw transcript dumps. Agent-written memories start as evidence-only pending review. Only user_confirmed or trusted imported memories can become instruction-grade.

Use Safe Agent Memory and Provenance as the operating guide for provenance, review status, use policy, and scope decisions.

Troubleshooting

Issue: agent-memory requires public.thoughts Solution: Run the core Open Brain setup first.

Issue: instruction-grade write fails Solution: This is usually correct. can_use_as_instruction is only allowed for user_confirmed or imported memory.

Issue: API cannot read tables Solution: Re-run the GRANT section at the bottom of schema.sql and redeploy the Edge Function so PostgREST reloads the schema cache.