Summary: MemPalace (52,000★, MIT) is an AI agent memory system that stores all conversational data raw instead of compressing it — achieving 96.6% on LongMemEval R@5 (the highest score ever recorded). It runs 100% locally, costs $0 in API fees, and can be installed into Claude Code in 2 commands.
The Memory Problem Nobody Solved — Until Now
Here's the thing about AI agents right now: they're goldfish.
You talk to Claude today, close the window, come back tomorrow — it has no idea what you talked about. Every session is a clean slate.
The industry has tried to fix this with:
- Summarization: Compress conversations into summaries → loses details
- Vector DBs: Store embeddings → loses exact recall
- Cloud APIs: OpenAI's memory API → you pay per token, data leaves your machine
All three have the same flaw: they throw information away.
Last month, a project called MemPalace hit GitHub and did something different. It stores everything — raw, unmodified, lossless — and relies on a hierarchical retrieval architecture to find exactly what you need, when you need it.
In 50 days, it got 52,000 stars. Here's why.
What Makes MemPalace Different?
Architecture: A Digital Palace
Palace (entire AI agent context)
├── Wing (project/topic category)
│ ├── Room (session/task)
│ │ ├── Drawer (individual message/command)
│ │ └── Drawer
│ └── Room
└── Wing
├── Room
└── Room
Four levels of hierarchy. Each level preserves the exact original content.
| Feature | MemPalace | OpenAI Memory | Summary-based |
|---|---|---|---|
| LongMemEval R@5 | 96.6% | ~72% | ~65% |
| Data location | 100% local | Cloud | Local |
| API cost | $0 | Pay per token | $0 |
| Lossless storage | ✅ Yes | ❌ Compressed | ❌ Summarized |
| License | MIT | Proprietary | Varies |
Why "Store Raw" Wins
Every other memory system summarizes conversations before saving. Here's why that's a problem:
User: "My server IP is 192.168.1.100, SSH key path is ~/.ssh/prod_key, and the deploy script is at /opt/deploy.sh"
Summary-based system saves: "User discussed server deployment."
MemPalace saves: The exact text, verbatim.
When you need to recall that SSH key path tomorrow, which system actually works?
How to Install (2 Commands)
pip install mempalace-mcp
# Add to your MCP config
# That's it. Your AI agent now has long-term memory.
It uses ChromaDB under the hood for vector storage, and the entire thing runs on your local machine. No cloud API, no data exfiltration, no monthly bill.
Where This Changes Everything
1. Code Projects
Start a project today, pick it up next week — your AI remembers every function signature, every bug you found, every decision you made.
2. Research
Read 50 papers in a week. Three weeks later, ask "What was the latency benchmark for that LLM from paper #37?" — and get the exact answer.
3. Personal Assistant
Your AI assistant remembers your preferences, your ongoing tasks, your family members' names. It gets better over time instead of resetting.
The Catch
It's still early. The GitHub is currently experiencing connectivity issues (peaked at 52K stars and the repo got hammered). The docs are minimal. You'll need to be comfortable with Python and MCP to set it up right.
But for an MIT-licensed project that already beats every commercial alternative on objective benchmarks? That's a trade I'd make.
Have you tried giving your AI agent long-term memory? Drop a comment below — I'd love to hear what works (and what doesn't).
If you found this useful, follow me for daily open-source deep dives. I cover AI tools, developer productivity, and the projects smart engineers are actually using.
Top comments (0)