Cognee

Cognee

by Topoteretes UG (Berlin)

Knowledge Engine for AI Agent Memory in 6 lines of code

See GitHub | Docs

Current state (2026-10-02)

• Open source, Apache 2.0; about 31.3k GitHub stars (the earlier 14.2k figure was outdated); v1.6.2 on PyPI (2026-09-29).
• Core operations are now documented as remember, recall, improve, forget (the older add / cognify / memify / search API names are still the lower-level pipeline vocabulary; check docs.cognee.ai for current signatures).
• Stores: graph database (traditional multi-store or a single-Postgres demo mode), vector database, Redis/session cache, relational metadata.
• Interfaces: Python, TypeScript and Rust SDKs, CLI, REST API, MCP server (Claude Code, Codex, Cursor, Cline), plus managed Cognee Cloud.
• Runs locally without API keys using small models (README claim).
• Funding: EUR 7.5M seed closed 2026-02-19, led by Pebblebed with 42CAP (eu-startups, Trending Topics; secondary sources; reportedly).
• Benchmarks: README reports BEAM conversational-memory scores of 0.79 (100K tokens) and 0.67 (10M tokens) with a noted reproduction gap; the research paper is Markovic et al. 2025, arXiv:2505.24478. The older “~93% vs ~60% for plain RAG” claim could not be verified and was removed.

What it is for

Persistent, evolving agent memory built as a knowledge graph: an LLM extracts entities and relationships from documents, code and conversations, and retrieval can traverse them (multi-hop) beyond plain vector search. Compare agent-memory-systems (Mem0, Zep/Graphiti, Letta) and graphrag.

Using with Obsidian

In the author’s view Cognee and Obsidian are complementary (untested here) — Obsidian is your human-readable PKM vault, Cognee adds a queryable knowledge graph on top of it for AI agents.

One possible approach (untested): Cognee MCP + mcp-obsidian

Run both MCP servers together so an AI assistant can:

1. Read/write your Obsidian vault via mcp-obsidian (pointed at your vault folder)
2. Feed notes into Cognee via cognee MCP to build a knowledge graph
3. Query with multi-hop reasoning across your entire vault

# Point cognee at your Obsidian vault  
import cognee  
  
await cognee.add("/path/to/ObsidianVault", dataset_name="obsidian")  
await cognee.cognify()  # builds graph from all markdown notes  
  
results = await cognee.search("what connects zettelkasten to my AI research?")  

What you gain over plain Obsidian graph view:

• Entity extraction across notes (not just backlinks)
• Semantic search + relationship traversal in one query
• Cognee documents an improve operation for refining the graph; the effect of querying on graph weights is not verified
• Agent-accessible: Claude / any MCP client can query your vault knowledge

Practical setup:

• Use Obsidian for writing and linking notes normally
• Run a nightly cognify job to keep the graph fresh
• Use Claude Code with both MCP servers to query and update the vault (untested here)
• See: Obsidian Forum thread on Cognee

Compared with Mem0 and QMD (qualitative)

• Mem0 extracts facts from conversations; its graph memory is on the paid Pro tier (mem0.ai/pricing, 2026-10-02). Cognee’s graph is free in the open-source version.
• QMD (local markdown search: BM25 + vector + reranking) adds no graph. The earlier “92% recall” figure for QMD is unverified and was dropped.
• Practical pattern (author’s suggestion, not benchmarked): QMD for fast vault search, Cognee for relationship queries.

Plans

• Open source: free, self-hosted (Apache 2.0).
• Cognee Cloud: usage-based, billed per tokens processed; per Cognee’s billing docs (opened 2026-10-05) billing is usage-based on tokens processed against prepaid credit; the first workspace is free (includes a token allowance, no payment method needed) and additional workspaces carry a monthly fee. No storage overage fee. Current prices: see cognee.ai.

Open items

• Seed funding from secondary press (eu-startups returned 403).
• Obsidian workflow and code snippet use the older add/cognify/search API and the Obsidian forum thread was not re-opened; test against v1.6 docs.
• Earlier claims dropped as unverified: ~93% multi-hop accuracy, 1M+ pipelines/month, Bayer and dltHub as users, 6-stage cognify details.

Sources

• https://github.com/topoteretes/cognee (accessed 2026-10-02)
• https://pypi.org/project/cognee/ (accessed 2026-10-02)
• https://www.eu-startups.com/2026/02/german-ai-infrastructure-startup-cognee-lands-e7-5-million-to-scale-enterprise-grade-memory-technology/ (search snippet, 2026-10-02)
• https://docs.cognee.ai/cognee-cloud/functionality/account-and-billing (opened 2026-10-05)
• https://mem0.ai/pricing (2026-10-02)