Cognee vs MemPalace
Two of the top memory, side by side: score, setup, license, activity and what each review found.
Cognee
Memory engine that turns documents and code into a knowledge graph
MemPalace
Local verbatim memory for coding agents on ChromaDB with 45 MCP tools
| What we compare | Cognee | MemPalace |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 64, popular | 87, widely used |
| Freshness | 100, active | 100, active |
| Maintenance | 82, healthy | 84, healthy |
| Easy to run | 67, easy | 50, easy |
| Agent-ready | 85, ready | 70, partly |
| Facts from GitHub and the README | ||
| Stars | 32.1k | 59.5k |
| License | Apache-2.0 (permissive) | MIT (permissive) |
| Last commit | Oct 2026 | Oct 2026 |
| Last release | Oct 2026 | Sep 2026 |
| Language | Python | Python |
| Docker | Yes | Yes |
| GPU | Not needed | Not needed |
| arm64 or Apple Silicon | Mentioned | Mentioned |
Cognee
Cognee builds persistent agent memory by extracting entities, relationships and chunks from text, code and sessions into a graph and vector index with hybrid recall. Without an LLM key it uses local GLiNER extraction and embeddings; adding a key enables generated answers via OpenAI, Ollama or other providers. It runs as a library, CLI, REST API (port 8000), UI (3000) and MCP server (8001).
Who it is for: Developers building graph-backed memory for agents and company knowledge
Strengths
- Keyless mode: local GLiNER extraction and embeddings, no cloud LLM required
- Claude Code and Codex plugins, OpenClaw plugin, MCP server, Python, TypeScript and Rust SDKs
- Imports memory from Mem0, Letta, Zep or Graphiti via the COGX format
- Apache-2.0; research paper and BEAM evaluation published
Weaknesses
- Single-Postgres graph store is a demo; production version is a licensed product
- API defaults to multi-tenant mode; local use needs ENABLE_BACKEND_ACCESS_CONTROL=false
- Bundled GLiNER extractor is a demo; higher-accuracy version requires contacting the vendor
- UI launcher needs Node.js/npm and Docker for its MCP service
- no GPU
- Docker + Compose
- Compose runs Neo4j, PostgreSQL, Redis
- Models: local GLiNER + embeddings (keyless), OpenAI, Ollama, other providers per docs
- port 8000
MemPalace
MemPalace stores conversation history verbatim, never summarised, and retrieves it by scoped semantic search over a palace of wings (people, projects), rooms (topics) and drawers. It runs locally with Python 3.9+ and ChromaDB by default, exposes 45 MCP tools plus a CLI, mines Claude Code, Codex and Cursor transcripts via hooks, and needs no API key: 96.6% R@5 on LongMemEval without an LLM.
Who it is for: Coding-agent users who want local, inspectable, verbatim session memory
Strengths
- Verbatim storage; nothing is summarised or paraphrased
- 96.6% R@5 on LongMemEval with no LLM or API key; results reproducible from the repo
- Pluggable backends: ChromaDB, sqlite, Rust-native, Milvus, Qdrant, pgvector
- Multi-arch Docker image; auto-save hooks for Claude Code, Codex and Cursor
Weaknesses
- First run downloads an 80 to 300 MB embedding model; Docker needs network then
- No native Android/Termux; GPU image is x86_64-only and unpublished
- Docker image runs as uid 1000, so bind mounts must be readable by that uid
- README warns about impostor domains distributing malware
- no GPU
- Docker + Compose
- Models: local embeddings (MiniLM, EmbeddingGemma), OpenAI-compatible embedding endpoints, Ollama
More in Memory
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.