Free resource
MemPalace: Local AI Memory That Stops Claude From Forgetting
MemPalace is a local-first AI memory system that stores your conversations word for word and brings the right context back into your next Claude Code session. It hits 96.6% recall at 5 results on LongMemEval with no LLM and no API key, and nothing leaves your machine unless you opt in. Here is what is inside and how to install it.
What it does
MemPalace is a local-first AI memory system. It stores your conversation history as verbatim text, indexes it in a structured palace (people and projects become wings, topics become rooms, content lives in drawers), and returns the right context back to your next session via semantic search.
Nothing leaves your machine unless you opt in.
Why it matters
Every Claude Code session resets. You spend the first 10 prompts re-explaining the project, the constraints, the past decisions. MemPalace fixes that.
It hits 96.6% recall at 5 results on LongMemEval with no LLM and no API key, just semantic search and a clean index. The hybrid pipeline pushes that to 98.4% on a held-out set. Numbers are reproducible from the repo.
What's inside
- Verbatim conversation storage, no summarization or paraphrasing
- Structured index: wings (projects/people), rooms (topics), drawers (content)
- Pluggable retrieval backend, default is ChromaDB
- 29 MCP tools for palace reads/writes, knowledge-graph operations, agent diaries
- Temporal entity-relationship graph with validity windows, backed by SQLite
- Two Claude Code hooks for periodic and pre-compaction auto-save
- Per-message recall via
mempalace sweepon top of file-level chunks
Install
Requires Python 3.9+, a vector store backend (ChromaDB by default), and about 300 MB of disk for the default embedding model.
pip install mempalace
mempalace init ~/projects/yourapp
mempalace mine ~/projects/yourapp
mempalace search "why did we switch to GraphQL"
For Claude Code, Gemini CLI, MCP-compatible tools, and local model wiring, see the docs site.