A knowledge graph your agents can search.
A governed memory across AIs and people.
Cognee is a mature, open-source memory layer that turns documents, code and conversations into a self-hosted knowledge graph, with an MCP server and coding-agent plugins. Kemory starts from the other end: it captures your AI chats automatically, recalls them in any AI and shares memory between people under deny-by-default rules and an audit log.
Last verified 7 November 2026 against primary sources
Two good answers.
Different jobs.
- Your main job is turning documents, code or databases into a queryable knowledge graph.
- You want an Apache 2.0 stack you can self-host today, with your choice of local models and storage.
- You are building agent memory in Python and want graph traversal, connectors and conflict resolution included.
- You want chats from ChatGPT, Claude, Gemini, Perplexity, Manus and Grok captured automatically and recalled in any AI.
- You need to share memory with other people, across organisations, with consent, credit, audit and revocation.
- You want governance on every plan: Gatekeeper, deny-by-default access, just-in-time consent and a hash-chained audit log.
See each one
doing its job.
Screenshots taken 7 November 2026. Cognee is a trademark of its owner, shown here to identify the product.
At a glance
The same rows on every comparison, so you can read one page against another.
*89.6% end-to-end answer accuracy on LongMemEval-S, open-weight Llama-3.3-70B reader, six runs. See the benchmarks or the full method on docs.sekondbrain.ai. A dash means not documented, not proven absent.
The same job,
done two ways.
Entity and relationship extraction, graph traversal and code graphs are Cognee’s core, and it is a large, active Apache 2.0 project you can self-host today. It ships ready-made connectors for Google Drive, Gmail, Slack, Obsidian, Asana and OneNote, and can run with local models and no cloud LLM key.
The numbers,
and how each was run.
These rows use different datasets, splits, readers and retrieval budgets, so they cannot be compared directly. We publish the method, every run including the lower ones, and use an open-weight reader rather than a frontier model.
| Product | Benchmark | Score | What it measures | Setup | Reader model | Runs |
|---|---|---|---|---|---|---|
| Kemory | LongMemEval‑S | 89.6% | End-to-end answer accuracy | 500 questions, oracle split | Llama-3.3-70B (open-weight) | Six, with variance |
| Cognee | BEAM 100K | 0.79 | LLM-judged, self-reported | 20 questions, one conversation | Not stated | Four rounds |
| Cognee | BEAM 10M | 0.67 | LLM-judged, self-reported (exploratory) | Synthetic long-context conversations | Not stated | Not stated |
What changed at Cognee
Products move quickly. These are the dated changes this comparison has been checked against.
- 18 September 2026v1.6.0: build and search text memory with local models and no cloud LLM key.
- 24 September 2026v1.6.1: Google Drive and Gmail connectors.
- 23 October 2026v1.6.6rc1: Obsidian, Asana and OneNote connectors; history-deletion controls.
- 29 October 2026v1.6.6: fact-conflict resolution API, with review on by default.
- 6 November 2026v1.6.7: tighter ingestion safety (a breaking change).
Specialists do one job well.
Kemory holds the thread across them all.
Every year brings better models and you will keep trying them. Kemory sits above the models rather than inside any one of them, so what you have worked out goes with you.
Tell one AI something once, and every AI you use just knows it. Start a thought in one and carry on in another, mid-sentence.
Every memory is permissioned and every change is provable. We do not train on your data. Delete anything, and prove it is gone.
Compaction runs on an open-weight model, so the frontier model only ever sees the compacted version — not the pile.
Already using Cognee? You do not have to replace it to try Kemory. Kemory connects to the AIs you use over MCP and Kora for Chrome.
Choosing Cognee instead?
Good — it fits you better right now.
Come back to Kemory when cross-AI, governed memory is what you need.
Five questions,
Kemory vs Cognee.
- What is the difference between Kemory and Cognee?
- Cognee turns your data into a self-hosted knowledge graph that agents can query. Kemory captures your AI conversations across ChatGPT, Claude, Gemini and others, recalls them in any AI over MCP, and lets you share memory with other people under consent and audit.
- Can I use Cognee with Claude Code or Cursor?
- Yes. Cognee documents an MCP server for Cursor, Cline and other MCP clients, and memory plugins for Claude Code, Codex and OpenClaw. Kemory offers MCP recall in Cursor, Codex, VS Code, Windsurf, Gemini CLI and more, and captures Claude Code sessions through its plugin.
- How are Cognee and Kemory priced?
- Cognee’s open-source core is free; Cognee Cloud is offered as a managed service, but we could not verify its prices on 7 November 2026. Kemory is free for 2,000 memories and two AIs, with Pro at $19.99 a month, Team at $30 a seat and Business at $55 a seat (10-seat minimum).
- Is Cognee open source?
- Yes. The core is Apache 2.0 and self-hostable with Docker Compose. Its README notes that Postgres as a graph store is a demo feature, with the production version sold as a licensed product. Kemory’s Apache 2.0 Community Edition is coming; enterprise self-host, your own cloud and air-gapped deployment are available now.
- Is Cognee’s BEAM score comparable with Kemory’s 89.6%?
- No. Cognee reports an LLM-judged 0.79 on BEAM at 100K tokens, from 20 questions on one synthetic conversation over four rounds. Kemory reports 89.6% end-to-end answer accuracy on the LongMemEval-S oracle split with an open-weight Llama-3.3-70B reader, across six runs. Different datasets, metrics and readers.
More comparisons
Credit where it’s due. The real difference, either way.