What 7 ms unlocks

When a database starts in 7 ms, it becomes a function call.

In-memory cold start: 7 ms. Durable local mode: 411 ms. At those numbers, the database stops being something you connect to and starts being something you invoke — born, used, and killed inside one agent execution.

The ephemeral context-graph lifecycle

A graph is born when the agent starts, dies when the agent finishes. Snapshot in between if you want to keep it.

Six things speed unlocks

Cold start isn't a vanity benchmark. It's the constraint that decides whether the database can be a real-time component of reasoning or an external service the agent has to wait on.

Context is scoped to a run.

Most of what an agent needs is relevant now, for this task, for this user. There's no global state to protect, no schema migration to coordinate, no shared cache to invalidate.

Real-time, not batch.

7 ms cold start plus sub-millisecond local queries means you can build, mutate, and tear down the context graph inside the reasoning loop. Mid-thought is fast enough.

Snapshots are first-class.

cognodb-dump writes plain JSON Lines. Version it in git, replay it for evals, branch it for what-if reasoning, ship it to another agent.

Zero standing footprint.

Between runs, nothing runs. You pay for memory and CPU only while the agent is actually thinking. Idle cost is literally zero.

Per-task isolation, for free.

Spin up a fresh graph per task. No cross-contamination between users, tenants, or runs. No mutex on a shared context.

Replayable by construction.

Because the graph is small and the cold start is negligible, you can re-run an entire reasoning session deterministically from a snapshot.

A different mental model

Traditional databases are services: long-lived, shared, slow to start, expensive to isolate. You connect to them. You manage their lifecycle separately from the workloads that use them.

CognoDB is closer to a library that happens to speak Bolt. The lifecycle of the graph matches the lifecycle of the task. When the task is over, the bytes are gone — unless you chose to snapshot them.

That changes how you architect agent systems. Instead of one giant context store with elaborate tenancy, you get one graph per agent run, cleaned up automatically.

Why graphs

Storage backends