MemPalace/mempalace

A local-first AI memory system that stores conversation and project history as verbatim text and retrieves it with semantic search, with no API calls required for its core path.

What it is

MemPalace mines projects and conversations into a searchable store and retrieves them with semantic search, without summarizing, extracting, or paraphrasing the original content. It is aimed at developers using AI coding agents — Claude Code, Codex CLI, Cursor, Gemini CLI — who want persistent, searchable memory that stays on their own machine. The index is structured: people and projects become "wings," topics become "rooms," and original content lives in "drawers," so searches can be scoped rather than run against a flat corpus. The retrieval layer is pluggable, defined by a backend interface, with ChromaDB as the default.

Key features

Verbatim storage with semantic retrieval — no summarization, extraction, or paraphrasing of stored content.
Pluggable storage backends: chroma (default), sqlite_exact, milvus, qdrant, and pgvector, selectable via --backend, MEMPALACE_BACKEND, or config.json.
CLI workflow (mempalace mine, search, wake-up, sweep) for mining project files and Claude Code session transcripts.
MCP server exposing 36 tools covering palace reads/writes, knowledge-graph operations, cross-wing navigation, drawer management, and agent diaries.
A temporal entity-relationship knowledge graph with validity windows (add, query, invalidate, timeline) backed by local SQLite.
Auto-save hooks for Claude Code, Codex CLI, and Cursor IDE that persist periodically and before context compaction.

Tech stack

Python, requires-python >=3.9 (classifiers list through 3.14).
Core dependencies: chromadb>=1.5.4,<2, pyyaml>=6.0,<7, huggingface_hub>=0.20, tokenizers>=0.15, numpy>=1.24, python-dateutil>=2.8, and tomli on Python < 3.11.
Embedding models: embeddinggemma-300m (multilingual, default for new installs) or all-MiniLM-L6-v2 (English-only, ~30 MB); ~300 MB disk for the default model.
Optional extras: milvus (pymilvus), pgvector (psycopg), extract (striprtf, markitdown for PDF/DOCX/PPTX/XLSX/EPUB), and gpu/dml/coreml ONNX runtime variants.
Build backend hatchling; dev tooling includes pytest, pytest-cov, ruff 0.15.20, mypy, and hypothesis.
Docker images for CPU and GPU (CUDA) that run the CLI or MCP server, persisting under /data.

When to reach for it

You want a local-first memory layer for AI agents where data does not leave your machine unless you opt in.
You need exact, verbatim recall of past conversations rather than model-generated summaries.
You are mining Claude Code, Codex CLI, or Cursor sessions and want them searchable and scoped by project or topic.
You want to wire memory into an MCP-compatible client through a stdio server.
You want to run entirely without an API key for the core retrieval path.

When not to reach for it

You want a hosted, managed memory service rather than something you install and run yourself.
Your workflow depends on summarized or extracted memories; MemPalace deliberately stores raw text.
You cannot spare roughly 300 MB of disk for the default embedding model.
You need a large-scale distributed vector store as a turnkey SaaS; server backends (Milvus, Qdrant, pgvector) are opt-in and self-configured.

Maturity signal

The project reports itself as version 3.6.0 with a "Development Status :: 4 - Beta" classifier, is MIT-licensed, and shows recent activity (last push in mid-2026 against an early-2026 creation date). The combination reads as young but under active development — the feature surface (multiple backends, an MCP server, a knowledge graph, auto-save hooks across three agents) is broad for its age, and the maintainers explicitly track corrections and public notices in a history file. Treat it as an actively developed beta rather than a settled 1.x release.

Alternatives

Mem0 — use it instead when you want a managed memory layer that extracts and summarizes rather than storing verbatim.
Zep — use it instead when you need a server-based memory service with its own end-to-end QA metrics.
Supermemory — use it instead when you prefer a hosted product over running and configuring a local store.

Notes

The README carries a prominent warning about impostor sites distributing malware and names its only official sources; that is unusual to see foregrounded in a project README.
The benchmarking stance is deliberately conservative: it publishes a "raw" 96.6% R@5 figure requiring no LLM, headlines a held-out 98.4% as the honest generalizable number, and declines side-by-side comparisons against Mem0, Mastra, Hindsight, Supermemory, or Zep on the grounds that mismatched metrics are not an honest comparison.
The LLM rerank path is described as model-agnostic and has been reproduced with Claude Haiku, Claude Sonnet, and a model via Ollama Cloud.

Wiki: MemPalace/mempalace