mnemory Persistent Memory
Give AI assistants searchable, personalized memory that carries across conversations.
- Source repo
- fpytloun/mnemory
- Stars
- ★ 284
- Last updated
- 3d ago
- License
- NOASSERTION
- Primary language
- Python
- FA score
- 63/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformChatGPT · Codex · Claude Code · OpenAI API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A Claude Code user who wants preferences, facts, and decisions to carry between sessions can configure mnemory as an MCP server.
- Not a fit if
- Users unwilling to provide an OpenAI-compatible API key
- Teams requiring verified erasure or compliance guarantees
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-17latest release v1.15.3 (2026-11-15)
- Source review
- 63/100 · Some gaps
What does this agent do, and when should you use it?
mnemory is a self-hostable MCP server that adds long-term memory to MCP-compatible assistants. Its unified LLM pipeline extracts facts, classifies metadata, deduplicates against existing memories, and resolves contradictions. Searchable summaries go into a vector store, while detailed material such as reports, code, and research is kept in an artifact store for retrieval on demand. Clients can use MCP tools or a REST API against the same backend, and a management UI provides browsing, semantic search, and relationship graphs. The quick start runs locally with uvx mnemory; production deployments can use Qdrant and S3/MinIO on infrastructure you operate.
Run uvx mnemory to start the service at http://localhost:8050/mcp; the default data directory is ~/.mnemory/. When storing information, mnemory calls an OpenAI-compatible API to extract facts, classify them, and check for duplicates or contradictions, then stores searchable content as vectors. Longer reports, code, or research can be held in the artifact store. For search, it generates multiple queries, runs them in parallel, and reranks results, with temporal awareness. An assistant can load pinned memories and recent context at the start of a conversation, then retrieve relevant memories as the conversation proceeds. The service also exposes memory CRUD, health checks, a REST API, a management UI, and a Prometheus /metrics endpoint.
- A Claude Code user who wants preferences, facts, and decisions to carry between sessions can configure mnemory as an MCP server.
- A ChatGPT user with the MCP connector can provide persistent personalized context to later conversations.
- Teams running Open WebUI or OpenClaw can use the documented integrations for automatic memory recall and storage.
- Teams looking up earlier decisions by date can use temporal-aware semantic search to retrieve relevant memories.
- Users keeping long research, reports, or code can store the detailed material as artifacts and retrieve it when needed.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
uvx mnemoryuvx --with "mcp<2" mnemory
- Ports
8050- Environment variables
OPENAI_API_KEY- Install notes cover
- Docker
How do you install or deploy this agent?
mnemory requires an OpenAI-compatible API key for LLM and embeddings; it automatically reads OPENAI_API_KEY. Start it with:
uvx mnemoryThe service defaults to http://localhost:8050/mcp and stores data in ~/.mnemory/. If an older uv on Alpine reports No module named 'mcp.server.fastmcp', the README gives this temporary command:
uvx --with "mcp<2" mnemoryThe README also documents Docker and pip availability, and production deployment with Qdrant and S3/MinIO, but the provided material does not include their exact installation commands.
How do you use this agent?
Configure an MCP client to connect to the local server. The README shows this Claude Code configuration:
{
"mcpServers": {
"mnemory": {
"type": "streamable-http",
"url": "http://localhost:8050/mcp",
"headers": {
"X-Agent-Id": "claude-code"
}
}
}
}Start a new conversation and the assistant can use memory through MCP tools. The REST API is another option; the supplied README excerpt does not include a request example.
What are this agent's strengths and limitations?
- One LLM pipeline handles fact extraction, metadata classification, deduplication, and contradiction resolution.
- Two-tier storage keeps searchable summaries separate from detailed artifacts such as reports and code.
- Multi-query semantic search, temporal awareness, and reranking help retrieve earlier decisions in natural language.
- MCP and REST API share a backend, with a memory management UI and Prometheus monitoring endpoint.
- LLM and embedding operations require an OpenAI-compatible API key; usage costs depend on the provider.
- Production deployments use components such as Qdrant and S3/MinIO, adding infrastructure configuration.
- Deletion is explicitly best effort: normal deletion retracts memories but preserves history, and privacy erasure is not guaranteed physical cleanup.
- The supplied README does not include Docker or pip install commands, so those deployment paths require consulting the deployment documentation.
How does this agent compare with similar options?
The README's LoCoMo benchmark table compares mnemory with Memobase, Mem0-Graph, Mem0, Zep, and LangMem. It reports an overall score of 73.2 for mnemory; Memobase scores 75.8, while the other listed systems score lower. These results apply to the stated configuration and benchmark, not every use case.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| mnemory Persistent Memory This agent | 63 · Some gaps | MCP serverFree + model costs | ★ 284 | 3d ago | Python | ChatGPT · Codex · Claude Code · OpenAI API |
| Athena | 48 · Major gaps | IDE extensionFree + model costs | ★ 647 | today | Python | Claude Code |
| LetsFG: AI-Native Flight & Hotel Search and Booking | 65 · Some gaps | CLIFreemium | ★ 2.3k | 5d ago | Python | Claude Code · OpenAI API · Claude API |
| Open Brain (OB1) | 55 · Major gaps | Self-hosted serviceFree + model costs | ★ 4.8k | 2mo ago | TypeScript | ChatGPT · Claude Code · Claude.ai |
How does FollowAgents rate this agent?
Why each dimension lost points
The materials describe local or remote vector storage, filesystem or S3 artifacts, LLM/embedding services, and authentication options. Delete-all is disabled by default, and ordinary retraction is distinguished from best-effort physical cleanup. Deductions: the quick start has no authentication by default and the server listens on 0.0.0.0; no per-action user confirmation is shown before capture; retention and external model processing of sensitive memories depend on operator configuration, while dependency security has version ranges but no audit evidence; publisher identity is unverified.
The docs say tools return JSON errors, log internal exceptions, and describe health checks, startup migrations, and service dependencies; alternative storage backends are documented. Deductions: tool counts differ (16 in README/AGENTS versus 19 in the server docstring); configuration calls gpt-5.4-mini the default while the benchmark assigns 73.2 to gpt-5-mini and says the newer model is expected to match or exceed it, leaving a static inconsistency; no execution evidence establishes operational reliability.
The materials address MCP clients, developers, and self-hosted production deployments, with multiple clients, Docker/pip/uvx, Qdrant/S3/Redis, and instruction modes. Deductions: capability boundaries are incomplete, and default proactive recall/capture may exceed user expectations; client behavior and the need for manual prompting vary by integration; although Python 3.11–3.13 and container architectures are covered, deployments still require compatible LLM, vector storage, and configuration.
Documentation is organized into quick start, configuration, architecture, API, deployment, and client guides, with concrete installation steps; both LICENSE and metadata identify Apache-2.0. Deductions: naming/tool counts vary; examples and FAQ coverage is useful but incomplete; some limitations are stated, including the lack of a verified erasure guarantee and model limitations, but coverage is not comprehensive; no changelog is in the sample, and maintenance responsibility is only partly indicated by author metadata and repository links.
The product offers fact extraction, deduplication, semantic search, two-tier memory, a management UI, and multiple interfaces, with MCP/REST responses and on-demand artifact retrieval. Deductions: value claims are mostly documentation assertions; the sample includes a benchmark table but not its reproduction materials; LLM/embedding calls for extraction and search introduce costs, latency, and operational work that are not quantified in depth.
Major README capabilities have corresponding implementation clues in configuration, the server entry point, plugins, and deployment docs; deletion limits, model selection, and test requirements are also stated. Deductions: the provided files cannot fully substantiate several benchmark and security/production-readiness claims; sources disagree on tool counts and model/benchmark descriptions; the docs distinguish measured from expected results, but this review did not run or independently reproduce the benchmark.
- Before deployment, configure authentication and restrict network exposure; evaluate the default proactive recall and capture behavior against client behavior and data policies.
- Deletion is explicitly best effort and is not a verified-erasure or compliance guarantee; benchmark results and production-readiness claims were not independently checked in this static review.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.15.3 Mnemory 1.15.3
Release notes excerpt
Fixes OpenClaw plugin discovery and temporal memory recall. Fixed - Declare plugin tools so OpenClaw can discover them. - Document how to replace native memory tools with Mnemory. - Verify the packaged plugin against supported OpenClaw versions. - Correct event-date and creation-date filters, include undated memories where appropriate, and preserve date bounds in generated search queries. Validation Unit tests: 1,655 passed, 20 skipped. Ruff, plugin packaging, Vitest and the OpenClaw Gateway compatibility matrix passed. Release E2E was not run for this …View on GitHub ↗ -
v1.15.2 Mnemory 1.15.2
Release notes excerpt
Fixed - Ship compiled JavaScript alongside the OpenClaw plugin sources, lock the build dependency, and verify that the packaged runtime loads before npm publication. Distribution - Publish multi-architecture container images to ghcr.io/fpytloun/mnemory, with source-repository labels and tested tag rules. - Document GHCR pull/build commands and retain the best-effort deletion warning alongside the Cognis introduction. The OpenClaw plugin remains at its separately published version 0.2.2. Deletion remains best effort, not a verified-erasure or compliance guarantee.View on GitHub ↗ -
openclaw-v0.2.2 OpenClaw Plugin v0.2.2
Release notes excerpt
Fixes npm installation with recent OpenClaw versions by shipping compiled JavaScript. Also fixes compatibility of the publishing verification with npm 12. See #16, #17, and #18.View on GitHub ↗
FAQ
Does using mnemory cost money?
Where is memory data stored?
~/.mnemory/. A production setup can use Qdrant for vectors and S3/MinIO for artifacts.Does deleting a memory guarantee verified physical erasure?
Do I need a particular assistant to use it?
Is mnemory Persistent Memory open source?
Is mnemory Persistent Memory safe to use?
Is mnemory Persistent Memory still maintained?
Who makes mnemory Persistent Memory?
Thanks — we use this to improve the page.