MemPalace Truth: A Feynman-Style Anatomy of a Memory System
> Forget the headline about "the Resident Evil actress's open-source AI memory system." That's a story. Stories are exciting, but stories are not understanding.
Start from a Concrete Picture
Imagine a notebook. Every time you chat with an AI, you copy the conversation into it word for word. No trimming, no summarizing—just the original text. Six months later, you want to know "why did we choose GraphQL over REST back then?"—you open the notebook and find that exact conversation. Not a summary, not "user prefers GraphQL," but the full discussion: the pain points with REST, GraphQL's advantages, and all the details you thought you'd remember later.
This is the core of MemPalace: store everything, then make it retrievable.
Other memory systems let the AI decide what's worth remembering. The AI sees "I like GraphQL" and saves "user prefers GraphQL," then throws away the original conversation. Sounds smart? The problem is the context of that preference is lost. Why do you like it? In what scenario? That's what you actually need.
The Memory Palace Is Not Just Marketing
MemPalace borrows its name from the ancient Greek mnemonic technique, but the data structure genuinely mirrors a palace:
| Level | Role | Analogy | |-------|------|---------| | Wing | One person or project | Library branch | | Room | A specific topic | A shelf in that branch | | Hall | Connects rooms, classifies by memory type | Corridor signage | | Closet | Summaries pointing to originals | Index cards | | Drawer | Original documents, verbatim | The original books |
Why this design? Because structure itself is the retrieval strategy.
They tested over 22,000 real conversation memories:
- Unstructured global search: 60.9% accuracy
- Restrict to one wing: 73.1% (+12%)
- Wing + hall: 84.8% (+24%)
- Wing + room: 94.8% (+34%)
- Local-first: you don't want conversations sent to the cloud
- Long-term companionship: an AI that remembers years of conversations
- Project memory: managing decision history and context across projects
- Cost-sensitive: Mem0/Zep charge; MemPalace is free
- Out-of-the-box: requires configuration, has a learning curve
- Enterprise support: no SLA, no professional customer service
- Large teams: currently better for individuals or small teams
Each structural layer acts as a search filter. Not magic—just shrinking the search space.
The 96.6% and 100% Story
MemPalace claims 96.6% (raw mode) and 100% (hybrid mode) on LongMemEval. The community erupted—"impossible," "marketing hype," "fake."
96.6% raw mode is credible
An independent developer reproduced the result on an M2 Ultra in under 5 minutes. Why? Verbatim storage has a natural advantage on retrieval tasks—nothing was lost at storage time, so retrieval finds it.
100% hybrid mode shows signs of optimization
The team admitted tuning against specific failing cases (from 99.4% to 100%), and used Haiku rerank (requiring an API call). A held-out test showed 98.4%—generalization isn't as perfect as the headline number.
AAAK compression was overhyped
The README once claimed "30x lossless compression." The community found:
1. It used a len(text)//3 heuristic token estimate, not a real tokenizer
2. In actual testing, AAAK mode scored only 84.2% on LongMemEval—12.4 points below raw mode's 96.6%
How did the team respond? They posted an honest update, admitted the issues, and fixed the README. That's what open source should look like—*brutal honest criticism makes open source work.*
Core Insight: Why MemPalace Works
Imagine finding a book. Most memory systems work like this: the librarian (AI) reads the book, writes a summary card, and throws the book away. Later you ask about a GraphQL book, and the librarian shows you the card. If the summary missed the part you need—too bad.
MemPalace's approach: all books stay on shelves (drawers), but the librarian built a detailed index system (wings/rooms/halls). When you ask, they consult the index, locate the shelf, and hand you the original book.
The key isn't compression—it's navigation.
Cargo Cult Detection
Is the palace structure a product or decoration?
The team says "+34% palace boost," but look closer—it's metadata filtering, a standard ChromaDB feature. Useful, but not revolutionary.
Is AAAK necessary or flashy?
Current evidence: raw mode 96.6%, AAAK mode 84.2%. AAAK doesn't even save tokens at small scale. It might help with large-scale repeated entities, but right now it looks like a "cool-looking" feature.
Is the 100% benchmark science or PR?
Both, per the team: 100% hybrid is real (with caveats), 96.6% raw is the honest number. But headlines say only 100%, omitting the asterisks. That edges into cargo cult science—the form is scientific (numbers, benchmarks), but the spirit is PR (picking the best number to show).
Is It Actually Useful?
Yes—in specific scenarios.
Good fits
Poor fits
Feynman-Style Summary
1. Store everything and retrieve vs. smart summarization: a real architectural trade-off. MemPalace chose the former—winning on retrieval accuracy, losing on storage efficiency. 2. Benchmark numbers can deceive: 96.6% and 100% are both "real," but in different contexts. The truly honest figure is 96.6% raw—zero API calls, zero external dependencies. 3. Open source's value lies in being questioned: the community found the AAAK problems and benchmark methodology flaws within 48 hours. The team responded honestly. That's how science should work. 4. The name and the code come from two people: Milla Jovovich initiated the architecture; Ben Sigman did the engineering. That's fine—software is collaborative—but know the difference.
A Final Warning
Don't believe because of star power, and don't dismiss because of the skeptics.
Try it yourself. pip install mempalace, import your conversations, search a question. See if you find what you want.
That's the way it is. Names aren't understanding, benchmarks aren't truth, star counts aren't quality. The only honest verification is your own usage.
This is the truth about MemPalace—an interesting project with real innovation and real over-marketing. Worth using, with your eyes open.
| Comparison | MemPalace | Mem0 | Zep | |-----------|-----------|------|-----| | LongMemEval R@5 | 96.6% (raw) / 100% (hybrid) | ~85% | ~85% | | Cost | Free | $19–249/mo | $25/mo+ | | Local deployment | Yes | No | Enterprise only | | API dependency | Zero (raw mode) | Yes | Yes | | GitHub Stars | 42K+ | 41K+ | – |
*Data sources: LongMemEval paper, official project documentation.*