### The highest-scoring AI memory system ever benchmarked. And it's free.
<br>
Every conversation you have with an AI — every decision, every debugging session, every architecture debate — disappears when the session ends. Six months of work, gone. You start over every time.
Other memory systems try to fix this by letting AI decide what's worth remembering. It extracts "user prefers Postgres" and throws away the conversation where you explained *why*. MemPalace takes a different approach: **store everything, then make it findable.**
**The Palace** — Ancient Greek orators memorized entire speeches by placing ideas in rooms of an imaginary building. Walk through the building, find the idea. MemPalace applies the same principle to AI memory: your conversations are organized into wings (people and projects), halls (types of memory), and rooms (specific ideas). No AI decides what matters — you keep every word, and the structure gives you a navigable map instead of a flat search index.
**Raw verbatim storage** — MemPalace stores your actual exchanges in ChromaDB without summarization or extraction. The 96.6% LongMemEval result comes from this raw mode. We don't burn an LLM to decide what's "worth remembering" — we keep everything and let semantic search find it.
**AAAK (experimental)** — A lossy abbreviation dialect for packing repeated entities into fewer tokens at scale. Readable by any LLM that reads text — Claude, GPT, Gemini, Llama, Mistral — no decoder needed. **AAAK is a separate compression layer, not the storage default**, and on the LongMemEval benchmark it currently regresses vs raw mode (84.2% vs 96.6%). We're iterating. See the [note above](#a-note-from-milla--ben--april-7-2026) for the honest status.
**Local, open, adaptable** — MemPalace runs entirely on your machine, on any data you have locally, without using any external API or services. It has been tested on conversations — but it can be adapted for different types of datastores. This is why we're open-sourcing it.
<sub>Reproducible — runners in <a href="benchmarks/">benchmarks/</a>. <a href="benchmarks/BENCHMARKS.md">Full results</a>. The 96.6% is from <b>raw verbatim mode</b>, not AAAK or rooms mode (those score lower — see <a href="#a-note-from-milla--ben--april-7-2026">note above</a>).</sub>
> The community caught real problems in this README within hours of launch and we want to address them directly.
>
> **What we got wrong:**
>
> - **The AAAK token example was incorrect.** We used a rough heuristic (`len(text)//3`) for token counts instead of an actual tokenizer. Real counts via OpenAI's tokenizer: the English example is 66 tokens, the AAAK example is 73. AAAK does not save tokens at small scales — it's designed for *repeated entities at scale*, and the README example was a bad demonstration of that. We're rewriting it.
>
> - **"30x lossless compression" was overstated.** AAAK is a lossy abbreviation system (entity codes, sentence truncation). Independent benchmarks show AAAK mode scores **84.2% R@5 vs raw mode's 96.6%** on LongMemEval — a 12.4 point regression. The honest framing is: AAAK is an experimental compression layer that trades fidelity for token density, and **the 96.6% headline number is from RAW mode, not AAAK**.
>
> - **"+34% palace boost" was misleading.** That number compares unfiltered search to wing+room metadata filtering. Metadata filtering is a standard ChromaDB feature, not a novel retrieval mechanism. Real and useful, but not a moat.
>
> - **"Contradiction detection"** exists as a separate utility (`fact_checker.py`) but is not currently wired into the knowledge graph operations as the README implied.
>
> - **"100% with Haiku rerank"** is real (we have the result files) but the rerank pipeline is not in the public benchmark scripts. We're adding it.
> - **96.6% R@5 on LongMemEval in raw mode**, on 500 questions, zero API calls — independently reproduced on M2 Ultra in under 5 minutes by [@gizmax](https://github.com/MemPalace/mempalace/issues/39).
> **Thank you to everyone who poked holes in this.** Brutal honest criticism is exactly what makes open source work, and it's what we asked for. Special thanks to [@panuhorsmalahti](https://github.com/MemPalace/mempalace/issues/43), [@lhl](https://github.com/MemPalace/mempalace/issues/27), [@gizmax](https://github.com/MemPalace/mempalace/issues/39), and everyone who filed an issue or a PR in the first 48 hours. We're listening, we're fixing, and we'd rather be right than impressive.
mempalace mine ~/chats/ --mode convos --extract general # general — classifies into decisions, milestones, problems
# Search anything you've ever discussed
mempalace search "why did we switch to GraphQL"
# Your AI remembers
mempalace status
```
Three mining modes: **projects** (code and docs), **convos** (conversation exports), and **general** (auto-classifies into decisions, preferences, milestones, problems, and emotional context). Everything stays on your machine.
After the one-time setup (install → init → mine), you don't run MemPalace commands manually. Your AI uses it for you. There are two ways, depending on which AI you use.
MemPalace also works natively with **Gemini CLI** (which handles the server and save hooks automatically) — see the [Gemini CLI Integration Guide](examples/gemini_cli_setup.md).
Decisions happen in conversations now. Not in docs. Not in Jira. In conversations with Claude, ChatGPT, Copilot. The reasoning, the tradeoffs, the "we tried X and it failed because Y" — all trapped in chat windows that evaporate when the session ends.
**Six months of daily AI use = 19.5 million tokens.** That's every decision, every debugging session, every architecture debate. Gone.
| Approach | Tokens loaded | Annual cost |
|----------|--------------|-------------|
| Paste everything | 19.5M — doesn't fit any context window | Impossible |
MemPalace loads ~600-900 tokens of critical facts on wake-up — your team, your projects, your preferences. Then searches only when needed. $10/year to remember everything vs $507/year for summaries that lose context.
The layout is fairly simple, though it took a long time to get there.
It starts with a **wing**. Every project, person, or topic you're filing gets its own wing in the palace.
Each wing has **rooms** connected to it, where information is divided into subjects that relate to that wing — so every room is a different element of what your project contains. Project ideas could be one room, employees could be another, financial statements another. There can be an endless number of rooms that split the wing into sections. The MemPalace install detects these for you automatically, and of course you can personalize it any way you feel is right.
Every room has a **closet** connected to it, and here's where things get interesting. We've developed an AI language called **AAAK**. Don't ask — it's a whole story of its own. Your agent learns the AAAK shorthand every time it wakes up. Because AAAK is essentially English, but a very truncated version, your agent understands how to use it in seconds. It comes as part of the install, built into the MemPalace code. In our next update, we'll add AAAK directly to the closets, which will be a real game changer — the amount of info in the closets will be much bigger, but it will take up far less space and far less reading time for your agent.
Inside those closets are **drawers**, and those drawers are where your original files live. In this first version, we haven't used AAAK as a closet tool, but even so, the summaries have shown **96.6% recall** in all the benchmarks we've done across multiple benchmarking platforms. Once the closets use AAAK, searches will be even faster while keeping every word exact. But even now, the closet approach has been a huge boon to how much info is stored in a small space — it's used to easily point your AI agent to the drawer where your original file lives. You never lose anything, and all this happens in seconds.
There are also **halls**, which connect rooms within a wing, and **tunnels**, which connect rooms from different wings to one another. So finding things becomes truly effortless — we've given the AI a clean and organized way to know where to start searching, without having to look through every keyword in huge folders.
You say what you're looking for and boom, it already knows which wing to go to. Just *that* in itself would have made a big difference. But this is beautiful, elegant, organic, and most importantly, efficient.
**Closets** — summaries that point to the original content. (In v3.0.0 these are plain-text summaries; AAAK-encoded closets are coming in a future update — see [Task #30](https://github.com/MemPalace/mempalace/issues/30).)
**Halls** are memory types — the same in every wing, acting as corridors:
-`hall_facts` — decisions made, choices locked in
-`hall_events` — sessions, milestones, debugging
-`hall_discoveries` — breakthroughs, new insights
-`hall_preferences` — habits, likes, opinions
-`hall_advice` — recommendations and solutions
**Rooms** are named ideas — `auth-migration`, `graphql-switch`, `ci-pipeline`. When the same room appears in different wings, it creates a **tunnel** — connecting the same topic across domains:
AAAK is a lossy abbreviation system — entity codes, structural markers, and sentence truncation — designed to pack repeated entities and relationships into fewer tokens at scale. It is **readable by any LLM that reads text** (Claude, GPT, Gemini, Llama, Mistral) without a decoder, so a local model can use it without any cloud dependency.
- **AAAK is lossy, not lossless.** It uses regex-based abbreviation, not reversible compression.
- **It does not save tokens at small scales.** Short text already tokenizes efficiently. AAAK overhead (codes, separators) costs more than it saves on a few sentences.
- **It can save tokens at scale** — in scenarios with many repeated entities (a team mentioned hundreds of times, the same project across thousands of sessions), the entity codes amortize.
- **AAAK currently regresses LongMemEval** vs raw verbatim retrieval (84.2% R@5 vs 96.6%). The 96.6% headline number is from **raw mode**, not AAAK mode.
- **The MemPalace storage default is raw verbatim text in ChromaDB** — that's where the benchmark wins come from. AAAK is a separate compression layer for context loading, not the storage format.
We're iterating on the dialect spec, adding a real tokenizer for stats, and exploring better break points for when to use it. Track progress in [Issue #43](https://github.com/MemPalace/mempalace/issues/43) and [#27](https://github.com/MemPalace/mempalace/issues/27).
A separate utility (`fact_checker.py`) can check assertions against entity facts. It's not currently called automatically by the knowledge graph operations — this is being fixed (track in [Issue #27](https://github.com/MemPalace/mempalace/issues/27)). When enabled it catches things like:
Create agents that focus on specific areas. Each agent gets its own wing and diary in the palace — not in your CLAUDE.md. Add 50 agents, your config stays the same size.
Each agent is a specialist lens on your data. The reviewer remembers every bug pattern it's seen. The architect remembers every design decision. The ops agent remembers every incident. They don't share a scratchpad — they each maintain their own memory.
**Optional auto-ingest:** Set the `MEMPAL_DIR` environment variable to a directory path and the hooks will automatically run `mempalace mine` on that directory during each save trigger (background on stop, synchronous on precompact).
