MemPalace Memory
Keep conversations and project material in a local, verbatim memory store and retrieve relevant context with semantic search.
- Source repo
- MemPalace/mempalace
- Stars
- ★ 60k
- Last updated
- today
- License
- MIT
- Primary language
- Python
- FA score
- 74/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeOpenAI API · Claude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A Claude Code developer who wants to import past sessions locally and search them in later sessions.
- Not a fit if
- Users who need a native Android Termux installation
- Teams requiring every retrieval and reranking step to avoid external model services
- Source review
- 74/100 · Some gaps
What does this agent do, and when should you use it?
MemPalace is a local-first AI memory system that stores conversation history and project files verbatim, without summarizing, extracting, or paraphrasing them. It uses ChromaDB by default for semantic retrieval and organizes people, projects, topics, and source material as wings, rooms, and drawers. Users can import and search content through the CLI, or let an MCP client read and manage memory through the MCP server; a Python API is also documented. Optional storage backends are available, and the system includes a temporal entity-relationship graph backed by local SQLite. The default retrieval and benchmark path needs no LLM or API key, though the first embedding operation downloads a model; optional LLM reranking or remote embedding adds a dependency on the chosen model or service.
Use mempalace mine to read project files or Claude Code sessions and store their original text in a palace. mempalace search finds relevant memories semantically, and mempalace wake-up loads context for a new session. Run mempalace audit to assess how the palace is organized. The MCP server exposes tools for palace reads and writes, knowledge-graph operations, cross-wing navigation, drawer management, agent diaries, and coordination. mempalace sweep can save conversations as one verbatim drawer per message; if LLM reranking is enabled, a model reranks retrieved candidates.
- A Claude Code developer who wants to import past sessions locally and search them in later sessions.
- An engineering team that needs to find decisions or technical discussions across projects, organized by project wing and topic room.
- A developer using an MCP client who wants an assistant to read or update a local memory store.
- An individual or team that wants a semantic index of project files and natural-language retrieval.
- A user who wants to maintain entity relationships with validity windows locally and query their timelines in SQLite.
How do you install or deploy this agent?
The recommended setup installs the CLI in an isolated environment with uv. Initialize a palace, then import project content. pipx is also supported; plain pip is recommended only inside an activated virtual environment. The first operation that needs embeddings downloads the selected model. For Docker, mount persistent storage at /data. Example MCP client configuration:
uv tool install mempalace
mempalace init ~/projects/myappDocker MCP server configuration:
{
"mcpServers": {
"mempalace": {
"command": "docker",
"args": [
"run", "-i", "--rm",
"-v", "mempalace-data:/data",
"-v", "/absolute/path/to/.claude/projects:/transcripts:ro",
"ghcr.io/mempalace/mempalace"
]
}
}
}How do you use this agent?
Import project files or Claude Code sessions, then search by question; use the wake-up command to load existing context into a new session. Replace the example paths with paths from your system. Example commands:
mempalace mine ~/projects/myapp
mempalace mine ~/.claude/projects/ --mode convos
mempalace search "why did we switch to GraphQL"
mempalace wake-up
mempalace auditWhat are this agent's strengths and limitations?
- Stores conversation text verbatim and supports semantic search without relying on summaries to preserve detail.
- ChromaDB runs locally by default; the README reports 96.6% raw R@5 on 500 LongMemEval questions without an LLM.
- Offers optional ChromaDB, sqlite_exact, rust_exact, Milvus, Qdrant, and pgvector storage backends.
- Includes an MCP tool set, CLI, Python API, and auto-save hooks for Claude Code and Codex CLI.
- Requires Python 3.10+; the embedding model must be downloaded, and the README lists model sizes of about 30 MB or 300 MB.
- Native Android Termux installation is unsupported; Android ARM64 users need a Debian PRoot container.
- Some backends need extra installation or service configuration; remote embeddings and LLM reranking depend on their configured services or models.
- On Linux, Docker bind mounts must be readable by the image's uid 1000; native installs also need local storage for the palace, configuration, and model files.
How does this agent compare with similar options?
The README identifies ChromaDB as the default and lists sqlite_exact, rust_exact, Milvus, Qdrant, and pgvector as optional backends. It says the project does not provide comparable benchmarks against Mem0, Mastra, Hindsight, Supermemory, or Zep, so it does not support a performance ranking against them.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| MemPalace Memory This agent | 74 · Some gaps | CLIFree + model costs | ★ 60k | today | Python | Codex · Claude Code |
| Remnic Agent Memory | 85 · Good | CLIFree + model costs | ★ 228 | 1d ago | TypeScript | ChatGPT · Codex · Claude Code · OpenAI API |
| Ori Mnemos | 80 · Good | CLIFree | ★ 331 | 19d ago | TypeScript | Claude Code · OpenAI API · Claude API |
| GraphMind Code Graph | 64 · Some gaps | Desktop appFreemium | ★ 216 | 5d ago | Rust | Claude Code · Claude.ai |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes local-first storage, optional remote embeddings, container mount boundaries, and read-only mining; weekly update checks are off by default and require authorization before an upgrade plan. However, the MCP server exposes many read, write, and coordination tools, and the supplied material does not show encryption or fine-grained access control for sensitive memories, so least privilege, sensitive-data handling, and external effects do not earn full marks. The security policy gives a private reporting path, and the README and authored-at document explain source-time fields and their limits. Publisher identity and operational ownership remain unclear, so they earn no extra credit.
The README, configuration, and CI files describe multiple backends, platform support, dependencies, and test gates. CI also documents a specific transient Chroma failure on Windows and limits retries to that error. Dependencies generally have version ranges rather than a fully pinned set, and failure-message evidence is mostly concrete examples in installation guidance, not proof of comprehensive runtime diagnostics; scores are therefore solid but below full marks.
The documentation covers developers, conversation and project-file memory, MCP, Docker, team deployments, multiple vector backends, and several platforms. It identifies opt-in dependencies, fallback behavior, GPU image architecture limits, and the Claude auto-save prerequisite. The three skills have described purposes and setup paths, but their contents are not included, so trigger precision cannot be fully established.
The README is organized around installation, backends, quickstart, benchmarks, tools, and documentation, with extensive command and environment-variable examples. It clearly states limits such as Termux and ARM GPU support, and the MIT license text is complete. Version information can be cross-checked across files, and a changelog is linked but not supplied; FAQ coverage and maintenance ownership are also insufficiently evidenced, so these criteria do not receive full marks.
The CLI, MCP tools, structured memory, and search-result fields are directly usable for ordinary workflows. The README gives benchmark figures and reproduction entry points, and avoids comparing unlike metrics across projects. The claimed marginal value relies mainly on project statements and referenced result files; the supplied materials do not include those results or an independent comparison. Resource costs are partly described through model downloads and disk requirements, but not fully measured.
The README links benchmark claims to result files, reproduction commands, and methodology documentation. The authored-at document carefully distinguishes ingest time, inferred dates, and verified authorship, with examples and backfill instructions. The evidence chain is clear, but benchmark artifacts, the full changelog, and skill contents were not supplied, limiting cross-source corroboration.
- Benchmark figures come from the project README and its referenced repository materials; this static review did not run reproduction steps or independently verify results.
- The palace may contain sensitive conversations; the supplied files do not describe at-rest encryption or fine-grained access control. Remote embeddings, model downloads, and server backends change data flows, so review the configured boundaries.