Cognee

You connect sources first. Cognee ships integrations for Slack, Notion, Linear and Google Drive on the paid tiers, and code indexing for your repositories, so the material arrives without you writing ingestion glue. Content is processed into a graph: entities, relationships and the provenance of each, rather than chunks scored by cosine distance.

Agents then read from it. The SDK is the direct route for code you control, and an MCP server is the route for Claude Code, Codex and any other MCP client, which is how most people will reach it without writing an integration at all. Billing follows the tokens the memory layer processes rather than seats or calls, so a quiet month costs less than a busy one, and workspaces are a separate fixed cost on top.

If you would rather not use the hosted service, the same engine runs locally or on your own stack under Apache-2.0, and the pricing page says so plainly rather than burying it.

• •Graph memory with an ontology - Builds entities and relationships over ingested content instead of a flat embedding index, which is what makes relational questions answerable rather than merely retrievable.
• •Source integrations - Slack, Notion, Linear, Google Drive and repository code indexing on the Standard tier, so the memory is built from where the work already lives.
• •Agent-native access - An SDK for your own code and an MCP server for Claude Code, Codex and other MCP clients, so an agent can use it without a custom integration.
• •Bi-temporal memory and conflict resolution - On Enterprise, the store tracks when a fact was true and when it was learned, and reconciles contradictions rather than storing both and hoping.
• •Provenance on answers - Enterprise recall cites where a claim came from, which is the difference between a memory layer you can audit and one you have to trust.
• •Self-hosting under Apache-2.0 - The full engine runs on your own infrastructure for nothing, with no feature gate on the open source build beyond the hosted conveniences.

Teams building agents that need continuity across sessions, and engineering organisations where the answer to a question is genuinely spread across a Linear ticket, a Slack thread and a pull request. It suits people already working through MCP clients, because the integration cost there is close to zero. The usage-priced Standard tier fits workloads that are spiky rather than steady, and the Apache-2.0 self-hosted build fits anyone who cannot send company content to a third party at all.

It is a weaker fit if you want a simple key-value memory for a single agent, where Mem0 or a plain vector store will be less machinery for the same result, and a reader who does not already know the difference will not see it from the marketing. Be clear-eyed about the Enterprise tier too: it is not a plan you buy but a fixed-scope consulting engagement with your engineers and theirs, offered in six, twelve and twenty-four month lengths, which is a heavier commitment than the Start building button on the homepage suggests.