One memory, every agent.

Stage-1 extraction. One structured row per run, generated from raw events. This is what makes messy agent work portable across harnesses.

create table run_summaries (
  run_id        text primary key references agent_runs,
  summary       text not null,
  extracted     text not null,         -- candidate memory
  keywords      text[],
  entities      jsonb,
  decisions     jsonb,
  artifacts     jsonb,
  generated_at  timestamptz not null
);

Promoted, owned memory. Durable facts the consolidation layer decided to keep — each scoped, sourced, and supersedable.

create table memory_items (
  memory_id     text primary key,
  scope         text not null,         -- global|project|client|agent
  topic         text not null,
  content       text not null,
  source_runs   text[] not null,
  source_agent  text,
  confidence    real,
  superseded_by text,                  -- evolve, don't delete
  expires_at    timestamptz
);

Provenance spine — our edge over the Codex snapshot. Every claim carries a status and its evidence, so weak guesses can never harden into facts.

create table claims (
  claim_id    text primary key,
  memory_id   text references memory_items,
  status      text not null,           -- fact|inference|unverified|refuted
  confidence  real,
  source_type text,                    -- person|profile|jon|record
  source_ref  text,
  verified_by text
);

The reliability spine. Lease + watermark + retry for every background worker — extract, consolidate, re-embed, expire. Crashes resume; nothing is lost in silence.

create table jobs (
  kind          text not null,         -- extract | consolidate | embed
  job_key       text not null,
  status        text not null,         -- pending|running|done|error
  worker_id     text,
  lease_until   timestamptz,
  retry_at      timestamptz,
  retry_left    int not null,
  input_watermark      timestamptz,
  last_success_mark    timestamptz,
  primary key (kind, job_key)
);

Recall across everything. A pgvector index over promoted memory — ask in natural language, get the right item regardless of which agent wrote it.

create table embeddings (
  memory_id  text references memory_items,
  embedding  vector(1536),
  model      text
);
create index on embeddings
  using ivfflat (embedding vector_cosine_ops);