Here is a thing that happened last week and you probably still do not believe it: Milla Jovovich (Alice from Resident Evil, Leeloo from The Fifth Element ) pushed an open-source AI memory system to GitHub, claimed a perfect score on the industry benchmark, watched it hit 15,000 stars in days, and then got publicly fact-checked by the AI developer community within 48 hours.
Real project. Disputed benchmarks. And a genuinely useful tool that European bootstrappers can install for free today.
That combination is worth your time.
TL;DR: MemPalace is a real, working, MIT-licensed open-source AI memory system built by Milla Jovovich and developer Ben Sigman. The code is on GitHub, it runs locally with zero API costs, and independent reviewers confirm the architecture is interesting and functional. The claimed benchmark scores, specifically the 100% on LongMemEval and LoCoMo, are contested due to methodology issues. The honest numbers (96.6% on LongMemEval in raw mode, 88.9% on LoCoMo without LLM reranking) still outperform paid competitors. For bootstrapping startups in Europe who cannot afford $500+/year in memory API costs, this is worth evaluating seriously right now.
Why This Story Hit Like a Rock Through a Window
The AI memory space was, until April 5, 2026, a quietly boring corner of developer infrastructure. You had Mem0 and Zep charging monthly fees to give your AI assistant persistent memory across sessions. Both hover around 85% on the standard LongMemEval benchmark. Paid. Cloud-dependent. Another subscription. Yikes.
Then a GitHub repository appeared under the username milla-jovovich. The launch tweet from co-creator Ben Sigman crossed 1.5 million impressions. The repo hit 7,000 stars within 48 hours and is now past 15,000.
The story wrote itself: actress outperforms VC-funded AI companies with a free tool built in her spare time. Tech Twitter lost its mind.
But here is what actually matters if you run a bootstrapped startup in Europe and you are trying to get more out of AI without spending more money: the architecture is genuinely different, the controversy is mostly about marketing, and the tool is free to use right now.
Let me break it down properly.
What Is MemPalace, Actually?
MemPalace is an open-source AI memory system that gives large language models persistent, cross-session memory. When you end a Claude or ChatGPT session today, everything disappears. Six months of daily AI conversations — decisions, debugging sessions, creative reasoning — gone. MemPalace solves that problem by storing all conversation data locally and using vector search to retrieve relevant memories on demand.
It runs entirely on your machine. Zero API costs. No cloud dependency. No subscription.
The architecture draws on the ancient Greek “method of loci” — the memory palace technique that orators used to memorize entire speeches by placing ideas in imagined rooms of a building. MemPalace applies the same logic to AI memory: your conversations get organized into wings (people and projects), halls (types of memory), and rooms (specific ideas).
Here is the technical stack:
- ChromaDB for vector storage and semantic search
- SQLite for a temporal knowledge graph that tracks facts with valid_from and valid_to dates
- AAAK compression — an experimental dialect that packs repeated entities into fewer tokens, designed for scale
- MCP integration with 19 tools, so Claude, ChatGPT, and other LLMs can query your memory automatically
The four-layer memory stack loads roughly 170 tokens of critical context on wake-up — your team, your projects, your preferences — then searches only when needed. Compare that to six months of daily AI use producing around 19.5 million tokens of lost context. The math is not complicated.
Who Actually Built This?
Jovovich designed the concept and core architecture. Developer Ben Sigman, CEO of Bitcoin lending platform Libre Labs, handled engineering. They built it using Anthropic’s Claude Code, an AI-assisted coding tool, over several months.
Jovovich announced the project via Instagram video on April 7, 2026, explaining she had spent months meticulously organizing AI files for a separate gaming project and kept running into the same problem: AI could not reliably retrieve information even when the files were well-organized. Reading about how ancient Greek orators memorized speeches pointed her toward the solution.
The repository is live at github.com/milla-jovovich/mempalace under her own account. The commit history is real. The code is real. The MIT license means you can use it, modify it, and build on it with no restrictions.
Sean Ren, a USC professor of computer science and CEO of Sahara AI, reviewed the approach and called it a general method for organizing information that could scale across AI frameworks. He also added the caveat that early benchmark results have not yet been validated in real-world deployments outside controlled tests. That is an honest assessment and you should keep it in mind.
The Benchmark Controversy: What X Got Right and What It Missed
People on X calling this fake are wrong about the project. They are closer to right about the numbers.
Here is the specific controversy, in plain language.
The LongMemEval claim: MemPalace claimed a 100% score. Independent reviewers, including a detailed analysis published by Penfield Labs, found that the 100% was achieved after targeted fixes for the three questions that were previously failing, plus LLM reranking. The X Community Note on Ben Sigman’s launch post states the held-out score is 98.4%. The raw verbatim mode without reranking scores 96.6%.
The LoCoMo claim: MemPalace claimed 100% with top_k=50. The problem: the LoCoMo dataset has conversations with 19 to 32 sessions each. Setting top_k=50 means you retrieve more sessions than actually exist, so the ground-truth session is always in the candidate pool regardless of what the embedding model does. At that setting, the retrieval step is effectively bypassed — you are just handing all sessions to Claude Sonnet and asking it to do reading comprehension. The honest LoCoMo number, also from their own benchmark documentation, is 88.9% with hybrid scoring and no LLM reranking.
The Mempalace team acknowledged these issues publicly and directly. Their README now includes a note thanking the community for brutal honest criticism and stating they would rather be right than impressive. That is a good sign for an open-source project.
The wider context: benchmark disputes are endemic in the AI memory space. Zep has accused Mem0 of inflated LoCoMo numbers. Mem0’s CTO has fired back with counter-analysis. Letta has published separate research questioning reproducibility on the same benchmarks. MemPalace walked into a benchmarking war that was already ugly before they arrived.
The honest summary: MemPalace scores 96.6% on LongMemEval in raw mode. That still beats Mem0 and Zep, which hover around 85%. And it does it for free, locally, with no API dependency.
What This Means for Bootstrapped Startups in Europe
I have been running bootstrapped startups in Europe since 2018 — first CADChain, which uses blockchain to protect IP for CAD and 3D models, then Fe/male Switch, a gamified startup education platform for women. Both companies run lean. Every tool has to justify its cost or it gets cut.
Here is the honest bootstrapper’s take on MemPalace.
The problem it solves is real. If you use Claude or ChatGPT daily for strategic decisions, product reasoning, content, customer communication, or code, you are losing context every single session. That loss compounds. You explain the same architecture decisions twice. You re-derive the same conclusions. You waste time that, as a bootstrapper, you cannot afford to waste.
The cost math is compelling. Paid memory tools can run $10 to $50 per month per user depending on usage. For a lean European team of four people, that is up to $2,400 per year. MemPalace costs nothing. The setup time is the only investment.
The privacy angle matters in Europe. GDPR is not optional. Sending your conversations to a third-party cloud memory service creates compliance questions that a bootstrapped team does not have time to answer properly. MemPalace keeps everything on your machine. That is not a small thing.
The risk is immaturity. The project launched six days ago. It will have rough edges. Installation friction. Missing Windows support. Limited documentation. If you need production-grade stability today, wait three months. If you can handle early-adopter friction, install it this week and start building your memory base now.
At Fe/male Switch, we track every AI decision that touches the game mechanics or the educational curriculum. The idea of having a local memory system that preserves that institutional knowledge across sessions, without paying a monthly fee and without sending our IP to a cloud service, is genuinely attractive.
MemPalace vs. Paid Competitors: The Honest Comparison
| Feature | MemPalace | Mem0 | Zep |
|---|---|---|---|
| Cost | Free (open-source) | Paid (SaaS) | Paid (SaaS) |
| Data location | Local (your machine) | Cloud | Cloud |
| LongMemEval score | 96.6% (raw mode) | ~85% | ~85% |
| Privacy / GDPR | Full local control | Vendor-dependent | Vendor-dependent |
| MCP integration | 19 tools | Varies | Varies |
| Setup complexity | Moderate (CLI) | Low (API key) | Low (API key) |
| Windows support | Not yet | Yes | Yes |
| Maturity | Days old | Established | Established |
| License | MIT | Commercial | Commercial |
The table does not lie. MemPalace wins on cost, privacy, and benchmark score. It loses on maturity and ease of setup. For a bootstrapped European startup that cares about both budget and data sovereignty, this trade-off is worth evaluating now.
How to Actually Use MemPalace: A Practical SOP
If you want to test it today, here is the setup process. This works on Mac and Linux. Windows support is listed as a feature request in the active issue tracker.
Step 1: Install
pip install mempalaceStep 2: Initialize your world
mempalace init ~/projects/myappThis sets up the palace structure for a specific project. Run it once per project you want to track.
Step 3: Mine your existing data
# For project files (code, docs, notes)
mempalace mine ~/projects/myapp
# For conversation exports (Claude, ChatGPT, Slack)
mempalace mine ~/chats/ --mode convosStep 4: Connect to Claude via MCP
claude mcp add mempalace -- python -m mempalace.mcp_serverAfter this, Claude automatically queries your memory when relevant. You do not type mempalace commands manually anymore.
Step 5: Search anything
mempalace search "why did we decide to use ChromaDB instead of Pinecone"The wake-up command loads your full context into any local model:
mempalace wake-up > context.txtPaste context.txt into your local model’s system prompt and it knows your history.
Insider Tricks for Bootstrappers
Start with your most expensive recurring AI task. If you spend two hours per week re-explaining context to an AI assistant, that is 100 hours per year. Mine those conversations first and test whether MemPalace retrieves the decisions that matter.
Use the temporal knowledge graph for business decisions. The SQLite-based knowledge graph tracks facts with validity windows — facts can expire when things change. This is genuinely useful for startups where team structures, product directions, and technical decisions shift frequently. Set valid_to dates on anything time-sensitive.
Export your Claude Code conversations first. MemPalace natively supports Claude Code JSONL exports, Claude.ai JSON, ChatGPT conversations.json, and Slack exports. If you use Claude Code daily, you already have months of mineable context sitting in your exports folder.
Do not run it on Windows yet. The community has flagged Windows compatibility as an open issue. Save yourself the friction and test on a Mac or Linux machine first.
Watch the benchmark numbers critically. The 96.6% raw mode score is the number to trust. The 100% claim requires LLM reranking which adds latency and complexity. For most startup use cases, raw mode is the right starting point.
Contribute to the project. MemPalace is MIT-licensed and the team is actively merging PRs. If you are a developer and you find an issue, filing it or fixing it makes the tool better for everyone and builds your profile in a fast-moving community.
The Mistake Everyone Is Making About This Story
The skeptics on X who called MemPalace a “publicity stunt” or “botslop” made one fundamental error: they confused benchmark controversy with project inauthenticity.
The code is real. The commits are real. The architectural ideas are real and independently validated by a USC computer science professor. Brian Roemmele, founder of The Zero-Human Company, deployed MemPalace to 79 employees within days of the public launch. That is not a publicity stunt deployment.
The benchmark marketing was aggressive. That is a separate issue from whether the tool works.
And here is the thing about benchmark wars in the AI memory space: they are universally dirty. Zep accused Mem0 of inflated numbers. Mem0 fired back. Letta published research questioning both. MemPalace arrived into a space where the credibility bar for benchmark claims was already on the floor.
The Mempalace team acknowledged the criticism directly and committed to fixing it. That is the right response. The community should hold them to it.
What Bootstrappers in Europe Should Do Right Now
Here is the practical breakdown of actions based on your current situation.
If you use Claude or ChatGPT daily: Install MemPalace on a Mac or Linux machine this week. Mine your last three months of exports. Run it for two weeks and track whether context retrieval saves you meaningful time. The cost of testing is zero.
If you are GDPR-conscious: This is one of the few AI memory tools where data sovereignty is built into the architecture, not bolted on as a feature. That alone is worth the setup friction for European teams.
If you are building AI products: The MCP integration with 19 tools is worth studying regardless of whether you use MemPalace in production. The approach to memory architecture — keep everything raw, index intelligently, let semantic search do the work — is a legitimate design philosophy that competes with the summarization-based approach used by Mem0 and Zep.
If you need Windows support: Wait. Check the GitHub issue tracker in a month. The community is active and Windows support is a top feature request.
If you are evaluating AI tools for your team’s budget: Add MemPalace to your short list. The honest benchmark numbers (96.6% LongMemEval) still beat paid tools. The cost is zero. The risk is early-stage immaturity, not fundamental architectural failure.
At Learn Dutch with AI, the use case is slightly different — conversation memory is the product, not just an operational convenience. Persistent memory that keeps every learning interaction, tracks vocabulary decisions, and retrieves past exercises without API costs is the kind of infrastructure that could meaningfully reduce operating costs for a language learning platform.
Similarly, for restaurants building their digital presence — healthy restaurants in Malta use AI tools to manage content, respond to reviews, and build SEO presence — a local memory system that tracks brand voice decisions, content history, and customer preference patterns across AI sessions could replace a paid tool with zero ongoing cost.
The pattern holds across different sectors. When you are bootstrapping, the question is always: what can I replace with something free that performs as well or better? MemPalace, at its honest benchmark numbers, makes that case.
The Bigger Lesson for European Bootstrappers
Milla Jovovich built a useful tool because she was frustrated with the existing options. She was a daily AI user who hit a wall, researched the problem, found a better approach, and built it. The fact that she is an actress is the headline. The fact that the tool works is the point.
That is the pattern European bootstrappers should recognize and replicate — not specifically in AI memory, but in their own domains.
You do not need to be a career developer to ship something useful with AI-assisted coding tools. Anthropic’s Claude Code is the tool Jovovich and Sigman used to build MemPalace. Fe/male Switch uses zero-code and low-code tools to build full gamified educational pipelines. CADChain embedded AI-assisted development early to stay lean while building IP protection infrastructure that would otherwise require a large engineering team.
The tools are better and cheaper than they have ever been. The barrier to shipping something real has never been lower. That is not a comfort — it is a warning. Your competitors are also aware of this.
The question is whether you act on it first.
Mistakes to Avoid When Adopting Open-Source AI Tools
Based on experience running bootstrapped startups in Europe for the past eight years, here are the traps to watch for.
Treating benchmark scores as ground truth. The LongMemEval controversy around MemPalace is a clear illustration: marketing benchmarks and real-world performance are different things. Always ask what the honest baseline number is and test it in your specific context.
Adopting early-stage tools for mission-critical workflows immediately. MemPalace is six days old at the time of writing. Use it for exploratory work and personal productivity first. Wait for community validation before putting it in a critical production workflow.
Ignoring the setup cost. Free tools are not always cheap. If MemPalace takes your team four hours to set up correctly, that is four hours of productive time. Calculate the actual cost of adoption, not just the license fee.
Missing the GDPR angle. If you are sending business conversations to a cloud memory service, check whether that service processes data in the EU or has adequate data processing agreements. MemPalace eliminates this question by design.
Skipping the community. The MemPalace community on X crossed 161 members within days. These are the people finding bugs, writing documentation, and building integrations. Joining the community is free and accelerates your learning curve dramatically.
FAQ
What exactly is MemPalace and what does it do?
MemPalace is an open-source AI memory system built by actress Milla Jovovich and developer Ben Sigman. It gives large language models like Claude and ChatGPT persistent memory across sessions by storing all conversation data locally on your machine. Instead of your AI starting fresh each session, MemPalace mines your past conversations and project files, organizes them into a navigable palace structure (wings, halls, rooms), and retrieves relevant context automatically via semantic search. It runs on ChromaDB and SQLite, requires no API key, and costs nothing. The project is MIT-licensed and available at github.com/milla-jovovich/mempalace.
Is MemPalace actually real or is it a marketing stunt?
MemPalace is a real project with real code, real commits, and a real GitHub repository under the username milla-jovovich. Independent technical reviewers have confirmed the architecture is functional and interesting. A USC computer science professor reviewed the approach and validated the general concept. Brian Roemmele deployed it to 79 employees within days of launch. The controversy is about benchmark marketing, not about whether the project exists or works. The tool is real. The claimed 100% benchmark scores are disputed. Those are two separate questions.
What are the real MemPalace benchmark scores?
The honest numbers, acknowledged by the team itself, are 96.6% on LongMemEval in raw verbatim mode (no LLM reranking), and 88.9% on LoCoMo R@10 with hybrid scoring and no LLM reranking. The 100% LongMemEval score required targeted fixes for specific failing questions plus LLM reranking, giving a held-out score of 98.4%. The 100% LoCoMo score used top_k=50 in a dataset where the maximum session count is 32, which means the retrieval step was effectively bypassed. Even the honest numbers outperform paid competitors Mem0 and Zep, which score approximately 85% on the same benchmarks.
How do I install and set up MemPalace?
Installation requires Python and pip. Run pip install mempalace to install. Then run mempalace init ~/your-project to set up the palace structure for a project. Mine existing data with mempalace mine ~/your-project for project files or mempalace mine ~/chats/ --mode convos for conversation exports from Claude, ChatGPT, or Slack. Connect to Claude via MCP with claude mcp add mempalace -- python -m mempalace.mcp_server. After that, Claude queries your memory automatically. Windows support is not yet available as of April 2026.
How does MemPalace compare to Mem0 and Zep for a bootstrapped startup?
For a bootstrapped startup in Europe, the comparison comes down to cost, privacy, and benchmark performance. MemPalace is free and MIT-licensed. Mem0 and Zep are paid SaaS products. MemPalace runs locally with full data sovereignty, which matters for GDPR compliance. Mem0 and Zep are cloud-based. MemPalace scores 96.6% on LongMemEval in raw mode; Mem0 and Zep score approximately 85%. MemPalace loses on maturity and ease of setup. The trade-off is worth evaluating for lean teams that cannot afford ongoing memory tool subscriptions and need to keep their data on-premise.
Did Milla Jovovich actually code MemPalace herself?
Jovovich designed the concept and core architecture. Developer Ben Sigman, CEO of Libre Labs, handled the engineering and implementation. They built the system together using Anthropic’s Claude Code over several months. Jovovich has publicly described her role as conceptual: she defined how AI memory should work based on her frustration as a daily AI user, and Sigman executed that vision in code. The GitHub commits are listed under both accounts. This is a common collaboration pattern in tech — founder vision plus technical execution — and it does not make the project any less real or functional.
What is the AAAK compression in MemPalace and why does it matter?
AAAK is an experimental lossy abbreviation dialect designed to pack repeated entities into fewer tokens at scale. The README initially claimed it saved tokens at small scales, but the team acknowledged this was incorrect after community feedback — AAAK does not save tokens for small datasets and the original README example used an inaccurate token count. At scale, with thousands of repeated entity references, AAAK is designed to reduce token usage significantly. The team has been transparent about correcting the initial claims. For a bootstrapped startup’s typical use case, the raw verbatim storage mode (which achieves the 96.6% benchmark score) is the appropriate starting point.
What does MemPalace mean for AI memory costs at an early-stage startup?
The cost impact is direct. Paid AI memory tools can run $10 to $50 per user per month. For a team of four using AI tools daily, that is $480 to $2,400 per year just for memory infrastructure. MemPalace costs nothing. The only investment is setup time. For a bootstrapped European startup already stretched across multiple tools and subscriptions, replacing a paid memory service with a free, locally-run, MIT-licensed alternative that scores higher on the benchmark is a straightforward financial decision once the tool matures past its first weeks.
Is MemPalace safe to use with sensitive business data?
Because MemPalace runs entirely locally — on your machine, using ChromaDB and SQLite, with no cloud dependency and no API key requirements — your data never leaves your infrastructure. That is a meaningful distinction from cloud-based memory services. For European startups subject to GDPR, local data storage eliminates the need for data processing agreements with memory service vendors and keeps all conversation history under your direct control. As with any open-source tool, review the code before using it with highly sensitive data, and check the current state of the repository as it evolves quickly.
What should I watch for as MemPalace develops over the next few months?
Watch for three things. First, independent real-world performance reports from teams that have deployed it in production for more than a few weeks — benchmark scores tell you one thing, daily usage tells you another. Second, Windows support, which is a top community feature request and will determine whether the tool is viable for teams not running Mac or Linux. Third, the trajectory of the project’s governance: open-source tools launched in a burst of viral energy sometimes lose momentum when the initial excitement fades. Track the commit frequency and community issue resolution rate over the next 60 days as a leading indicator of long-term health. The MeanCEO blog covers open-source AI tools and bootstrapper resources as they develop, and is worth following for ongoing updates.
