An open-source AI memory palace that broke the benchmark

MemPalace is an open-source, local-first memory system that gives any LLM persistent recall of your life, ongoing projects, and personal preferences without sending your data to the cloud.

Created by Ben Sigman and actress Milla Jovovich using Anthropic's Claude, the tool runs entirely on your local machine with zero subscriptions or API keys required.

How it works

Most current AI memory setups dump your conversation logs into a flat vector store or send them off to background agents in the cloud. MemPalace instead organizes your local chats into a navigable mental palace made of wings, halls, and rooms.

This structure sorts memories by domain and category rather than treating them as an endless list of facts. Before you even type your first prompt, your core context—family members, active work, and specific tastes—loads into the model in roughly 120 tokens.

To keep context overhead minimal, it relies on a proprietary AAAK compression technique that squeezes your broader life context down by 30x without losing fidelity. Any standard LLM can parse and read this native format out of the box.

The system also features built-in contradiction detection to catch mistaken names, swapped pronouns, or incorrect ages before they reach your screen. For semantic search across months of chats, the correct answer reliably lands in position one or two.

Benchmarks and setup

MemPalace posted historic numbers across the industry's standard memory benchmarks.

On LongMemEval, it logged the first perfect score on record with a 100% recall rate, answering 500 out of 500 questions correctly across every question category.

On the ConvoMem benchmark, it scored 92.9%, more than double the score of popular alternative Mem0. It also achieved a flawless 100% on LoCoMo, clearing complex multi-hop reasoning tasks including difficult temporal inference.

Setting it up is deliberately simple. The system requires only a single dependency, runs entirely on your hardware, and keeps personal logs strictly on-device. The project is completely open source under the MIT License on GitHub under milla-jovovich/mempalace.

Why it matters

If you care about local privacy and hate re-explaining your context every time you open a new chat window, MemPalace is well worth a spin right now. Squeezing your background down to roughly 120 tokens makes long-term memory practical without eating up your model context budget.

We think its local, structured palace approach is a smart rejection of bloated cloud vector architectures. That said, claiming 30x truly lossless compression on nuanced human memories is a bold promise; it remains to be seen how well the palace hierarchy holds up when thousands of chaotic, unstructured life details pile in over months of daily use.