MemPalace

https://github.com/MemPalace/mempalace

This is a fascinating crossover between Hollywood and the dev world. Milla Jovovich moving from slaying zombies to solving AI...

Open Url

Resident Evil Star Launches Open-Source AI Project on GitHub

You might not recognize the name Milla Jovovich immediately, but you’ve definitely seen her films. Whether she’s Alice in Resident Evil or Leeloo in The Fifth Element, she is a Hollywood hardcore action icon. However, in April 2026, she’s trending for a different reason: she released an open-source project on GitHub called MemPalace. An action star launching an open-source AI project? The contrast alone is worth the buzz. But how does the project actually hold up? To be frank, it has brilliant flashes of engineering mixed with some glaring flaws. 《Resident Evil Star Launches Open-Source AI Project on GitHub》

01 Why Did She Build This?

In the MISSION.md file, Milla explains her motivation. She has long used an AI assistant called Lumi for project collaboration, but she hit a wall that every heavy AI user hates: Context Compression. Once the context is compressed or cleared, the AI develops amnesia. She described how Lumi would wake up every session like a “new colleague,” asking what they should do today—even if they had been chatting for hours previously. After testing various memory systems on the market, she found them lacking.
“These systems are like empty warehouses,” she noted. “You pile in tons of information, but when you actually need something, you’re searching forever and coming up empty.”
She wanted a “fuzzy recall” capability—that feeling of “didn’t we talk about something like this before?”—which standard keyword searches simply can’t replicate. So, alongside engineer @bensig, she used Claude Code to build a solution born from real pain points. This wasn’t a PR stunt; it was a quest for a better tool.

02 What exactly is MemPalace?

In short: MemPalace is a locally-run AI long-term memory system. It saves your conversations and project files verbatim and retrieves them using semantic search. Key distinction: It stores everything word-for-word. No summaries, no rewrites. While most memory systems compress or summarize content to save space, MemPalace insists on the original text. The downside is higher storage costs; the upside is zero information loss. For personal use, disk space is cheap—information fidelity is expensive. The storage structure borrows from the “Method of Loci”:
  • Wing: The top-level unit (e.g., a specific project or person).
  • Room: Specific topics (e.g., auth-migration, deploy-process).
  • Closet: Compressed indices pointing to the original content.
  • Drawer: The verbatim raw conversation text.
  • Tunnel: Cross-wing associations. If the same topic appears in different projects, the graph layer automatically links them.
The inspiration comes from two “old-school” methods: the Ancient Greek Memory Palace technique and Niklas Luhmann’s Zettelkasten (slip-box) method. It sounds elegant, but as we’ll see, elegance doesn’t always equal performance.

03 The 4-Layer Memory Stack

This is the most valuable engineering contribution of the project. MemPalace utilizes a progressive 4-layer loading stack:
  1. L0 (Identity Layer): ~50 tokens. Tells the AI who it is and who it serves.
  2. L1 (Key Story Layer): 500–800 tokens. The system automatically picks the 15 most important moments from all memories (grouped by Rooms) as the core context for every wake-up.
  3. L2 (On-Demand Recall): 200–500 tokens. Only loads specific “Room” content when a relevant topic is mentioned.
  4. L3 (Deep Search): A full semantic retrieval across the entire “Palace.”
The core philosophy is efficiency: you don’t need to dump all memories into the prompt. By loading on demand, a typical AI “wake-up” only uses 600–900 tokens, which is significantly cheaper than brute-forcing the entire history.

04 The Hard Data (and the Hard Truth)

First, the good news: MemPalace’s raw semantic search is impressive. On the LongMemEval benchmark, the Raw mode achieved a 96.6% R@5 recall rate. This is achieved entirely locally using ChromaDB vector retrieval—no APIs, no LLMs required. The Hybrid v4 mode hit 98.4% on a held-out set (tested on 450 unseen entries), proving the retrieval strength is genuine.

The Fatal Flaw: The “Palace” is a Penalty

Here is the part people often miss. The project’s own BENCHMARKS.md reveals a troubling trend:
  • Raw Mode (Pure semantic search): 96.6%
  • With Rooms (Metadata filtering): 89.4% (A 7.2% drop)
  • With AAAK Compression: 84.2% (A 12.4% drop)
The core selling point—the elaborate “Palace” architecture of Wings, Rooms, and Closets—actually drags down retrieval performance. The 96.6% success rate is thanks to the underlying ChromaDB, not the palace structure. This happens because structured retrieval is harder than pure semantic search; adding filters often accidentally excludes the correct answer. While the structure helps organize information for a human, it currently hinders the AI’s accuracy.

05 Ecosystem and Integration

Despite the performance controversy, the integration is top-notch.
  • MCP Server: Includes 29 tools covering retrieval, knowledge graphs, and cross-wing navigation.
  • Compatibility: Works with Claude Code, Gemini CLI, and other MCP-compatible tools.
  • Cost Savings: Background hooks auto-archive conversations. Milla claims this saves roughly $1.13 per turn in repeated transmission costs.
  • Privacy: Everything stays local. It uses SQLite for knowledge graphs and ChromaDB for vectors. No API keys required for core functions.

The Controversy: A Lesson in Transparency

After going public on April 5, the project faced immediate backlash. Early marketing claimed the “Palace” architecture provided a +34% performance boost (it was actually negative) and a 100% recall on certain benchmarks (which turned out to bypass the retrieval process). To the team’s credit, they spent 10 days rewriting the README, retracting the claims, and publishing all raw data and 42 regression tests. It was a rocky start, but they chose accountability over doubling down.

Final Verdict: An Intriguing Experiment

MemPalace is more of a fascinating experiment than a production-ready tool. The Highlights: The 4-layer memory stack is a brilliant engineering approach to token management, and the local-first, privacy-centric stance is a breath of fresh air. The Lowlights: The “Palace” structure is currently a performance bottleneck, and the early hype was misleading. If you are a power user who loves tweaking your AI workflow and can tolerate some “beta” rough edges, it’s worth a try. If you need a stable, enterprise-grade memory solution, you might want to wait. Quick Start: 
pip install mempalace
mempalace init ~/projects/myapp
mempalace mine ~/projects/myapp
nothing