> Forget the "Resident Evil star open-sources an AI memory system" headline. That's a story. Stories are exciting, but stories are not understanding.
---
Start With a Concrete Picture
Imagine you have a notebook. Every time you chat with an AI, you copy the conversation verbatim into it. No trimming, no summarizing — just the original words. Six months later, you want to know "why did we choose GraphQL over REST back then?" You open the notebook and find that exact exchange. Not a summary, not "user prefers GraphQL," but the full discussion: the pain points of REST you tried, the advantages of GraphQL, and all the details you assumed you'd remember later.
This is the core of MemPalace: store everything, then make it retrievable.
How do other memory systems work? They let the AI decide what's worth remembering. The AI sees you say "I like GraphQL" and stores a line like "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.
---
Memory Palace Is Not a Marketing Term
MemPalace borrows its name from the ancient Greek mnemonic technique, but it's not a gimmick — the data structure really is a palace:
| Level | Purpose | Analogy | |------|------|------| | Wing | One person or one project | A branch library | | Room | A specific topic | A subject shelf in the branch | | Hall | Connects rooms, classifies memory types | Corridor signage | | Closet | Summaries pointing to raw content | Index cards | | Drawer | Original documents, unchanged | The original books |
Why this design? Because the structure itself is the retrieval strategy.
They tested over 22,000 real conversational memories:
- Global unscoped search: 60.9% accuracy
- Scoped 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 multiple projects
- Cost-sensitive: Mem0/Zep charge; MemPalace is free
- Out-of-the-box use: requires configuration, has a learning curve
- Enterprise support: no SLA, no professional support
- Large teams: currently better for individuals or small teams
Each layer of structure acts as another filter on the search. Not magic — just shrinking the search space.
---
The 96.6% and 100% Story
Now, the controversial benchmark.
MemPalace claims 96.6% (raw mode) and 100% (hybrid mode) on LongMemEval. The community immediately reacted — "impossible," "marketing gimmick," "fake."
Let me separate the claims:
96.6% raw mode is credible
An independent developer reproduced it on an M2 Ultra in under 5 minutes. Why? Because verbatim storage has a natural advantage on retrieval tasks — if you don't lose information at storage time, retrieval finds it.
100% hybrid mode shows signs of optimization
The team admitted they tuned against specific failing questions (going from 99.4% to 100%) and used Haiku rerank (requiring API calls). Held-out tests show 98.4%, meaning 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. Actual testing: AAAK mode scored only 84.2% on LongMemEval — 12.4 points below raw mode's 96.6%
How did the team handle the criticism? They published an honest update, acknowledged the issues, and corrected the README. That's how open source should work — *brutal honest criticism makes open source work.*
---
Core Insight: Why MemPalace Works
An analogy. 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 for "that book about GraphQL," and the librarian shows you the summary card. But if the summary doesn't mention the part you actually need? You're done.
MemPalace's approach: all books stay on the shelves (drawers), but the librarian builds a detailed index system (wings/rooms/halls). When you ask a question, the librarian doesn't recall from memory — they consult the index, find the exact shelf location, and hand you the original book.
The key isn't compression. It's navigation.
---
Cargo Cult Detection
Feynman-style sharp questions:
Is the palace structure a product or decoration?
The team touts a "+34% palace boost," but look closer — that's metadata filtering, a standard ChromaDB feature. Useful, but not revolutionary.
Is AAAK necessary or flashy?
Evidence so far: raw mode 96.6%, AAAK mode 84.2%. AAAK doesn't even save tokens at small scale. It might help at large scale with repeated entities, but right now it looks like a "looks cool" feature.
Is the 100% benchmark science or PR?
The team says both: 100% hybrid is real (with caveats), 96.6% raw is the honest number. But the headline says 100% without the asterisks. That's on the edge of cargo cult science — the form is scientific (numbers, benchmarks) but the spirit is PR (picking the best number to show).
---
So Is It Useful?
Yes. In specific scenarios, very.
Good fits for MemPalace
Bad fits
A Feynman-Style Summary
What MemPalace teaches us:
1. Store everything and retrieve vs. intelligent summarization is a real architectural trade-off. MemPalace chose the former, winning on retrieval accuracy and losing on storage efficiency.
2. Benchmark numbers can deceive. 96.6% and 100% are both "real," but in different contexts. The genuinely honest number is 96.6% raw — zero API calls, zero external dependencies.
3. The value of open source is being questioned. The community found AAAK's problems and benchmark methodology flaws within 48 hours. The team responded honestly. That's how science should work.
4. The name and the code were written by different people. Milla Jovovich initiated the architecture; Ben Sigman handled the engineering. That's fine — software is a collaborative art. But you should know the difference.
---
Finally, a Warning
Don't believe it because of star power, and don't dismiss it because of the skeptics.
Try it yourself. pip install mempalace, import your conversations, search a question. See if you find what you're looking for.
That's the way it is. Names are not understanding, benchmarks are not truth, star counts are not quality. The only honest verification is your own use.
---
This is the truth about MemPalace — an interesting project with real innovation and real over-marketing. Worth using, but with eyes open.
| Item | 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 | | API dependency | Zero (raw mode) | Yes | Yes | | GitHub Stars | 42K+ | 41K+ | - |
*Sources: LongMemEval paper, official project documentation*