🏛
Wings
Top-level containers that separate major domains — a wing for each project, person, or knowledge area. Wings keep unrelated memories from colliding.
MemPalace: The Open-Source AI Memory System That Scores 100%
Created by Milla Jovovich. First perfect score on the LongMemEval benchmark. MIT licensed. Runs 100% locally with zero API costs.
7,000+
GitHub Stars
100%
LongMemEval
MIT
Licensed
$0
API Cost
What Is MemPalace?
MemPalace is the first open-source AI memory system to score a perfect 100% on the LongMemEval benchmark, providing persistent cross-session memory for any large language model at zero cost.
MemPalace is an open-source AI memory system that gives large language models persistent, cross-session memory. When you close a conversation with ChatGPT or Claude, everything discussed is forgotten. MemPalace fixes that by storing every interaction and making it instantly retrievable in future sessions.
The key insight behind the MemPalace architecture is deceptively simple: store everything verbatim, then make it findable. Competing systems like Mem0 and Zep use an LLM to extract what the AI thinksis important and discard the rest. This inevitably loses context, nuance, and the user's original reasoning. MemPalace takes the opposite approach — it keeps every word and uses vector search to surface exactly what you need.
Under the hood, the MemPalace system uses ChromaDB for semantic vector search and SQLite for structured metadata storage. Both run entirely on your local machine — no cloud services, no API keys, no recurring costs. Install it with pip install mempalace and you have a fully functional AI memory system in under a minute.
MemPalace also introduces AAAK, a lossless compression dialect that achieves 30x compression on stored memories without losing any information. This means your memory palace can hold thousands of conversations while staying fast and lightweight on disk.
AI Memory Benchmark Comparison
How MemPalace performs against competing AI memory systems on the LongMemEval recall benchmark (R@5), cost, and deployment model.
| System | LongMemEval R@5 | API Required | Cost | Runs Locally |
|---|---|---|---|---|
| MemPalace (hybrid) | 100% | Optional | Free | Yes |
| MemPalace (raw) | 96.6% | None | Free | Yes |
| Supermemory ASMR | ~99% | Yes | Paid | No |
| Mastra | 94.87% | Yes (GPT) | API costs | No |
| Mem0 | ~85% | Yes | $19-249/mo | No |
| Zep | ~85% | Yes | $25/mo+ | No |
Benchmark data sourced from the LongMemEval benchmark paper. ConvoMem data from Salesforce Research. Last verified: April 2026.
As of April 2026, MemPalace is the highest-scoring AI memory system on all three major benchmarks: LongMemEval (100%), ConvoMem (92.9%), and LoCoMo (88.9%). No other system has achieved a perfect score on any of these benchmarks.
How It Works — The Memory Palace Architecture
MemPalace organizes AI memory using a spatial metaphor inspired by the ancient memory palace technique. Each level maps to a real storage concept.
🚪
Rooms
Each wing contains rooms for specific topics. A project wing might have rooms for architecture decisions, user research, and deployment notes.
🏰
Halls
Corridors that connect rooms by memory type: facts, events, discoveries, preferences, and advice. Halls make cross-room retrieval fast and intuitive.
🗄
Closets
Compressed summaries of conversations encoded in the AAAK dialect. Closets store 30x more information than raw text in the same token budget.
📂
Drawers
The original verbatim files — full conversation transcripts, code snippets, and reference material. Nothing is ever lost; drawers are the source of truth.
⚡
AAAK Compression
A 30x lossless compression dialect that MemPalace uses to pack dense information into minimal tokens. AAAK is what makes it possible to store thousands of sessions locally.
Who Created MemPalace?
Milla Jovovich — the actress known for the Resident Evil franchise — built MemPalace alongside developer Ben Sigman using Claude Code. The origin story is surprisingly practical: after months of intensive AI collaboration, Jovovich had accumulated thousands of decisions, ideas, and lines of reasoning across hundreds of conversations.
The frustration that sparked MemPalace was simple: existing AI memory systems decide what to remember for you. Mem0, Zep, and similar tools use an LLM to extract “key facts” and discard the rest. But Jovovich found that the discarded context — the reasoning, the alternatives considered, the nuance — was exactly what she needed most in future sessions.
The solution was MemPalace: store everything verbatim, compress it with the AAAK dialect for efficiency, and use vector search to surface the right memories at the right time. The result is the first AI memory system to score a perfect 100% on the LongMemEval benchmark — because when you keep everything, you never lose what matters.
Frequently Asked Questions About MemPalace
What is MemPalace?+
MemPalace is an open-source AI memory system that gives large language models persistent, cross-session memory. Unlike other systems that use AI to decide what to remember, MemPalace stores everything verbatim and makes it findable through a spatial metaphor inspired by the ancient memory palace technique. It uses ChromaDB for vector search and SQLite for structured storage, running entirely on your local machine.
Is MemPalace free?+
Yes, MemPalace is completely free and open-source under the MIT license. It runs locally on your machine with no API costs, no subscription fees, and no usage limits. The only optional cost is if you choose to use a cloud LLM for the hybrid search mode, but the raw vector search mode requires zero API keys.
How does MemPalace compare to Mem0?+
MemPalace scored 100% on the LongMemEval benchmark compared to Mem0's approximately 85%. The key architectural difference is that MemPalace stores all information verbatim and makes it findable, while Mem0 uses an LLM to extract and summarize what it considers important — which inevitably loses information. MemPalace is also free and runs locally, while Mem0 costs $19-249 per month and requires API access.
Does MemPalace require API keys?+
No. MemPalace's raw search mode works entirely locally with zero API keys. The optional hybrid mode can use an LLM API to enhance search results, but it is not required. The core system — ChromaDB vector storage, SQLite metadata, and the AAAK compression dialect — all run on your machine without any external dependencies.
What LLMs does MemPalace work with?+
MemPalace is LLM-agnostic. It works with Claude (via Claude Code or the API), ChatGPT, local models like Llama and Mistral, and any LLM that can read structured text. The memory system itself is a Python package that any LLM can interface with through its CLI or Python API.
How do I install MemPalace?+
Install MemPalace with a single command: pip install mempalace. It requires Python 3.10 or higher. After installation, run mempalace init to create your first memory palace. For a detailed walkthrough including Claude Code integration, visit our setup guide.
What is the best AI memory system in 2026?+
As of April 2026, MemPalace holds the highest scores on all major AI memory benchmarks: 100% on LongMemEval, 92.9% on ConvoMem, and 88.9% on LoCoMo. It is also the only system that runs entirely locally at zero cost. Other notable systems include Mem0 ($24M funded, ~85% LongMemEval), Zep (knowledge graph focus), and Letta (hierarchical memory).
How does MemPalace store memories?+
MemPalace stores all conversation data verbatim — it does not use an LLM to extract or summarize what it considers important. Instead, it keeps every word in ChromaDB (vector search) and SQLite (structured metadata), then compresses stored memories 30x using the AAAK lossless compression dialect. This approach preserves the full context, reasoning, and nuance of every conversation.
What is AAAK compression in MemPalace?+
AAAK is a custom lossless compression dialect developed for MemPalace that achieves approximately 30x compression on stored memories. Unlike lossy summarization used by other systems, AAAK preserves all original information in a shorthand format that any LLM can read natively without a decoder. For example, 1000 tokens of English text compress to roughly 30 tokens of AAAK while retaining complete fidelity.
Give Your AI Perfect Memory
MemPalace is free, open-source, and installs in under a minute. Join 7,000+ developers who chose the AI memory system with the highest benchmark score in history.