Agent memory that doesn't lie.
Durable state, tool-callable HTTP, and a reactive log of every action - all transactional on one substrate. Bring your own LLM key for the reasoning calls.
JSON-shaped rows with vector embeddings, queried together. The agent's view never lies about what was just written.
The DB's HTTP surface is the agent's tool surface - /sql, /vector, /graph, /ask. No SDK glue to maintain.
Every write is a WAL event. /watch turns the audit trail into a real-time stream - observable, replayable.
One agent watches another in real time.
A chain of specialist agents can hand tasks off via the substrate.
/watch on a prefix turns "agent B sees what
agent A just wrote" into an HTTP subscription, no message broker required.
State is durable. Replays are deterministic. Debugging a multi-agent system stops requiring time travel.
POST /v1/tenants/:t/batch
{
"ops": [
{ "sql": "INSERT INTO tasks ... RETURNING id" },
{ "vector": { "schema": "task_emb", "put": [...] } },
{ "sql": "INSERT INTO audit (action, agent) VALUES ('open', 'a-7')" }
]
}Reasoning calls use your provider, not ours.
The agent's LLM calls - via /ask or your own
stack - use the BYOK key you registered. No markup, no quota games, full audit in your provider's
dashboard. BYOK LLM details →
- not usMulti-agent orchestration primitives. We're infrastructure; pick your own framework on top.
- under designPersistent agent-state versioning + branchable replays.