What is MemPalace?
MemPalace is an open-source AI memory system that stores your conversations verbatim and makes them searchable via semantic search. Unlike tools like Mem0 or ChatGPT's built-in memory that use AI to extract and summarize (losing information in the process), MemPalace keeps every word of every conversation.
The result: 96.6% recall accuracy on the LongMemEval benchmark — nearly double what extraction-based tools achieve.
The Problem MemPalace Solves
If you use AI tools daily, you've probably experienced this:
- You had a detailed discussion about your architecture decisions 3 weeks ago — but you can't find it
- You keep re-explaining your tech stack, project context, or preferences to every new AI session
- Your ChatGPT conversation history is thousands of chats deep, and search barely works
Every AI tool today treats conversations as disposable. When the session ends, the context is gone. MemPalace fixes this by treating your conversations as a permanent, searchable knowledge base.
How MemPalace Works
MemPalace uses a spatial metaphor inspired by the Method of Loci (memory palace technique):
- Wings — Top-level categories (like "Engineering", "Marketing", "Design")
- Rooms — Specific topics within a wing (like "Auth Architecture", "SEO Strategy")
- Drawers — Individual conversation fragments stored with semantic embeddings
When you search for something like "why did we choose JWT?", MemPalace uses vector similarity search (not keyword matching) to find the most relevant conversation fragments — even if you never used those exact words.
The Technology
- Embeddings: all-MiniLM-L6-v2 (384-dimensional vectors)
- Vector database: ChromaDB (CLI version) or pgvector (web version)
- Search: Cosine similarity for semantic matching
- Integration: MCP (Model Context Protocol) for connecting to any AI tool
Two Ways to Use MemPalace
1. CLI Version (Self-Hosted)
The original open-source tool runs entirely on your machine:
pip install mempalace
mempalace mine ~/chats/ --mode convos
mempalace search "why did we choose GraphQL"- Pros: Fully local, data never leaves your machine
- Cons: Requires Python 3.9+, ChromaDB, terminal knowledge
- GitHub Repository →
2. Web Version (mempalace.me)
Upload your conversations in your browser, get a hosted MCP endpoint:
- Export your ChatGPT data (Settings → Data Controls → Export)
- Upload
conversations.jsonat mempalace.me - Get your MCP URL and connect any AI tool
- Pros: No Python, no terminal, 60-second setup
- Cons: Data is stored on our server (encrypted, token-protected)
- Try it now →
Benchmarks
MemPalace has been tested on established academic benchmarks:
| Benchmark | MemPalace | Mem0 | Zep | ChatGPT Memory |
|---|---|---|---|---|
| LongMemEval (R@5) | 96.6% | 49-85% | 63.8% | N/A |
| ConvoMem | 92.9% | 30-45% | — | — |
Benchmarks from LongMemEval (ICLR 2025) and ConvoMem (Salesforce). MemPalace scores come from raw verbatim storage mode with default ChromaDB embeddings.
What is MCP?
Model Context Protocol (MCP) is an open standard that lets AI tools access external data sources. Think of it like a USB port for AI — any MCP-compatible tool can connect to your memory palace with a single URL.
MemPalace provides MCP tools for:
- search — Find conversations by meaning
- list_wings — Browse your palace structure
- list_rooms — See conversations in a topic
- traverse — Read a full conversation
- status — Palace overview
Getting Started
The fastest way to try MemPalace is the web version. Upload your ChatGPT export, preview your palace, and connect your AI tools — all in 60 seconds.
For the full local experience, install the CLI version:
pip install mempalace
mempalace mine ~/Downloads/chatgpt-export/ --mode convos
mempalace search "what was our database decision"FAQ
How do I export my ChatGPT data?
Go to ChatGPT → Settings → Data Controls → Export data. You'll receive an email with a download link containing conversations.json.
Does it work with Claude conversations? The CLI version supports Claude exports. Web version support is coming soon.
Is it really free? The CLI version is MIT-licensed and completely free. The web version (mempalace.me) offers a free tier for personal use.
How accurate is the search? 96.6% recall on LongMemEval. This comes from storing raw text with vector embeddings rather than AI-extracted summaries. The key insight: keeping everything beats being selective.