Imagine this: you spend an entire week with Claude Code debugging a complex distributed systems problem. You find the root cause, define the architecture, write the fix. Then you close the session.
The following week, you open a new conversation. Your AI has no idea that any of that happened. You’re starting from scratch.
Every decision. Every architecture debate. Every “we chose Postgres over MySQL because…”. Gone.
MemPalace is an open-source project that exploded in recent weeks (42.9k stars and counting) with a simple but radical idea: instead of letting the AI decide what’s worth remembering, save everything — and make it searchable.
The Problem with “Smart” Memory
Most AI memory systems work like this: at the end of a conversation, an LLM reads it in full and extracts what it considers important. “The user prefers tabs over spaces.” “The project uses React 18.” Concise, token-efficient, neat.
But there’s a problem: the AI throws away the why. You lose the conversation where you explained why you chose React, why you moved away from REST, why you wrote that weird workaround. Without that context, the extracted preference is just noise.
MemPalace takes the opposite approach. It saves your actual conversations verbatim — the complete exchange, word for word — in a local ChromaDB instance. No summarizing. No AI deciding what matters. When you need something, semantic search finds it.
The Palace: a Mental Map for Your Conversations
The name comes from an ancient Greek memory technique: the Method of Loci, or memory palace. Speakers would memorize entire speeches by mentally placing ideas in rooms of an imaginary building, then walk through it to retrieve them.
MemPalace applies the same structure to your conversations with AI:
- Wings — high-level context (people you work with, projects)
- Halls — types of memory (decisions, architecture debates, debugging sessions)
- Rooms — specific ideas or conversations within a hall
This gives you a navigable map instead of a flat search index. You’re not just doing keyword search — you’re walking through a structured space that reflects how you actually work.
The Numbers (With Honest Context)
MemPalace claims the highest score ever published on LongMemEval: 96.6% R@5 in raw mode, tested independently across 500 questions, zero API calls.
It’s a real result — and it was independently reproduced by community members on an M2 Ultra in under 5 minutes.
But there’s something you need to know: that score is specifically for raw verbatim mode. The project also has an experimental compression feature called AAAK — and the founders were transparent about overstating its capabilities on launch day. AAAK mode currently scores 84.2% R@5, a regression of 12.4 points. The initial claim of “30x compression with no loss” was also incorrect (it’s lossy, and token savings only appear at scale, not in the small examples they showed).
Why mention it? Because the founders published a public correction note on April 7, acknowledging every issue in detail. In the open-source world, that kind of honesty is worth highlighting. The 96.6% result in raw mode holds. The rest is in progress.
Quick Start
pip install mempalace
# Initialize your "world" — projects, collaborators
mempalace init ~/projects/myapp
# Feed it your conversations
mempalace mine ~/projects/myapp # code, docs, notes
mempalace mine ~/chats/ --mode convos # Claude, ChatGPT, Slack exports
mempalace mine ~/chats/ --mode convos --extract general # classifies decisions, milestones, issues
# Search for anything you've ever discussed
mempalace search "why we migrated to GraphQL"
Your Claude and ChatGPT conversation exports go straight in. Slack ones too.
Integration with Claude Code and Codex
The repo includes both a .claude-plugin folder and a .codex-plugin folder — meaning it integrates directly into Claude Code and OpenAI Codex CLI workflows. It also comes with a full MCP server, so you can connect it to any MCP-compatible setup.
It’s genuinely model-agnostic. Claude, ChatGPT, Gemini, Llama, Mistral — if it reads text, it can work with MemPalace output.
Malware Warning: There Is No Official MemPalace Website
Community members have reported fake MemPalace websites appearing online, including some distributing malware. The project has no official website. The only legitimate source is the GitHub repository. If you find it anywhere else claiming to be official, don’t install anything.
Is It Worth Using?
MemPalace represents a significant philosophical shift from how most memory systems work, and the raw-mode benchmark results are real. The codebase is young — there are known open issues (a segfault on macOS ARM64, a shell injection in hooks, ChromaDB version pinning) — but the team is actively addressing them.
If you do a lot of agentic work and find yourself re-explaining the same context every session, MemPalace gives you a structured way to stop doing that. Local, free, open-source, no cloud required.
The palace is open. Go in and explore it.