Every conversation you have with an AI disappears when the session ends. Six months of architecture decisions, debugging sessions, and hard-won preferences - gone. You start over every time. We built Mempalace to fix that permanently.
Other memory systems solve forgetting by letting AI decide what matters. It extracts "user prefers Postgres" and discards the conversation where you explained why - the reasoning that makes the preference meaningful. Mempalace takes a fundamentally different approach: store everything, structure it intelligently, make it all findable. The judgment call about what matters belongs to you, not the model.
The method of loci (memory palace) is a mnemonic technique used since ancient Greece. Orators memorized hour-long speeches by placing ideas in rooms of a building they knew well. Walk through the building, find the idea. Mempalace applies this architecture to AI memory: every conversation has a place, and that place makes it findable.
Your conversations contain your best thinking - architecture decisions, personal preferences, private context. That data should live on your machine, under your control, permanently. Mempalace runs entirely locally. No cloud required. No account needed for the core product. Your memory is yours.
AAAK compression is structured text with a universal grammar. It works with any model that can read - Claude, GPT, Gemini, Llama, Mistral, anything. Mempalace is not a wrapper around one provider. It is infrastructure that works regardless of which AI you use today or tomorrow.
After months of deep work with Claude and GPT, patterns emerged. Every new session started from scratch. The AI didn't know about last month's architecture decision. It didn't know about the naming convention debate. It didn't know you'd already tried and rejected three approaches to the auth problem. Six months of context, gone.
Cloud memory products summarize your conversations, losing the verbatim context that makes retrieval meaningful. They decide what matters before you've asked the question. They require subscriptions and external APIs. None of them run locally. None of them work with every model. The gap was too large to ignore.
The palace structure came from reading about how memory actually works in humans - spatially organized, contextually anchored. AAAK came from the constraint that local models have tight context windows - you need compression that preserves everything without requiring a decoder ring. The benchmark result (96.6%) validated the approach.
The right approach for infrastructure is open. Privacy guarantees mean nothing if you can't inspect the code. Local-first means nothing if the tool is a black box. Mempalace is MIT licensed, fully open, and built to be extended. The benchmark runners are in the repo so you can verify the numbers yourself.
You cannot predict which context will be relevant in a future conversation. The answer is to store all of it and build retrieval that finds the right piece. Filtering before storage is lossy by definition.
Random storage and semantic search alone give roughly 62% recall. Adding palace structure - wings, halls, rooms, tunnels - pushes that to 96.6%. The metaphor is not cosmetic. It is load-bearing.
30x compression sounds like summarization. AAAK is not summarization - it is a dialect with a defined grammar that can be decoded losslessly. Every entity, emotion marker, and relationship survives compression intact.
The core product requires no external API. ChromaDB runs locally. SQLite runs locally. The MCP server runs locally. You can use Mempalace on an airgapped machine with a local Llama model and it works identically.
Every benchmark claim in the README is reproducible. The runners are in the repo. The dataset is public (LongMemEval). Run it yourself and verify. We don't make performance claims we cannot back with code.
Mempalace was tested on conversations - but the architecture generalizes. Code repositories, documentation, notes, emails. If it's text and you've thought about it, it can live in the palace.
Five commands from zero to an AI with perfect memory of your entire working history. Free and open source.