# About OceanDB

> OceanDB is sovereign, self-managing memory for AI agents, built by Ocean Labs.

Updated: 2026-07-11

OceanDB is sovereign, self-managing memory for AI agents. It gives the agents you already use one durable memory, then keeps that memory readable, sourced, and under your control.

## Why it exists

Memory is the part of an AI that should be most yours. Instead, it is often tied to one model, one product, or one conversation—and still asks you to repeat, summarize, and reorganize what already happened.

OceanDB asks for neither your labor nor your surrender. Agents capture durable context through MCP. The Dreamer tends it later, turning a raw log into a cited wiki and surfacing changes that need human judgment.

## What makes it different

- **One memory across agents.** Claude, ChatGPT, Cursor, Codex, and other MCP clients can use the same store.
- **A source you can inspect.** Consolidated pages cite the raw entries they came from.
- **Maintenance you control.** The Dreamer runs from your own agent with your chosen model.
- **Boundaries below the prompt.** Postgres row-level security and restricted database grants enforce access.
- **A complete way out.** Export the memory as Markdown or SQL.

## Ocean Labs

Ocean Labs builds OceanDB. The work begins with a simple position: memory should answer to the person or team it describes.

The product is built in public language even where the internals are technical. Its schema, agent contract, maintenance playbook, privacy model, and operating limits are documented so the memory does not become a hidden authority.

## How it is built

OceanDB uses Postgres for durable state, pgvector for semantic retrieval, and MCP for agent access. The data model separates raw entries, maintained pages, typed links, projects, tags, workspaces, and agent-authored insights.

The hosted server contains no consolidation model. The Dreamer is a client-side process: it reads the public playbook and works through the user's own authorized agent.

Read [what OceanDB is](/docs/introduction), inspect [the schema](/docs/schema), or [connect an agent](/docs/connect).