Mnemosyne

Data Sovereignty by Design

·Not a policy. Not a promise. A property of the architecture.

·527KB. Zero dependencies. Runs anywhere.

·No telemetry. No API keys. No data leaves your machine.

·Every memory is a Markdown file — readable, diffable, yours forever.

·We don't ask for permission to remember your data. We ask for none.

💡 What It Is & Where It Fits

Mnemosyne is a zero-dependency, local-first memory engine purpose‑built for AI agents. It integrates natively with OpenClaw, Hermes, and any agentic framework — and runs out‑of‑the‑box on Windows, Linux, and macOS.All this, in just 527KB — small enough to fit on a floppy disk, run on a Raspberry Pi, or embed into any edge device. No cloud, no bloat, just memory.

Wherever you need persistent, sovereign memory without the cloud:

• 🤖 Agentic workflows — give your agents long‑term recall without a single API call
• 🧠 Personal AI assistants — remember user preferences, conversation history, and past interactions, all stored locally
• 🛠️ Development tools — maintain context across multi‑hour coding sessions, even offline
• 📚 Learning companions — implement spaced repetition and review schedules, never exposing learner data to third parties

Whether you're deploying on a Windows workstation, a Linux server, or a macOS laptop, Mnemosyne works identically — because memory sovereignty shouldn't depend on your operating system.

🎯Why it is special

Mnemosyne doesn't use neural networks. It uses 140 years of cognitive psychology — from Ebbinghaus' forgetting curve (1885) to SAM compound-cue theory (1981), from the Zeigarnik effect (1927) to TEPA memory revocation (arXiv 2026). Thirty-one papers, ten retrieval factors, all encoded as pure mathematical formulas. It stores like a machine and recalls like a human — without a single API call.

Our compound-cue formula isn't fitted from training data — it's a direct translation of the SAM model into code. Five base weights (imp, recency, keyword, hitFreq, layerW), five cognitive biases (topic coherence, Zeigarnik, Primacy, Context, Testing Boost), and three post-processing filters (superseded, MMR, RIF). Every single factor traces back to a named researcher and a published year. Ask "why 0.35" — the answer is Raaijmakers & Shiffrin, 1981. Not a hyperparameter sweep.

Most AI memory systems are black boxes — feed everything to a large model, let it memorize, let it retrieve. Expensive, slow, and unreliable. We took a different path: 140 years of scientific research on how human memory actually works. We turned psychology formulas into code. No AI models, no API keys — yet the agent remembers like a person. Important things stick. Old memories fade but never disappear. A single keyword triggers entire conversations. That's not artificial intelligence. That's human intelligence, reverse-engineered.

🖥️ Demo / Screenshots

Special Thanks

桦染霜（Tiktok ID）

⚡ Quick Start

git clone https://github.com/ElonAug7/Mnemosyne-agentmemory-engine-openclaw-hermes
cd Mnemosyne-agentmemory-engine-openclaw-hermes/Mnemosyne-v6.4
bash install.sh
open http://localhost:8765

🚀 Latest: v6.4 — Beats every embedding system, without a single embedding

On our Memory-Native Evaluation benchmark (80 queries, 11 systems including Mem0, LlamaIndex, qwen-agent, Google ADK):

System	nDCG@10
Mnemosyne v6.2	0.046
raw BM25 baseline	0.185
embedding systems (Mem0 / LlamaIndex / ...)	0.12–0.16
Mnemosyne v6.3+	0.238 — 5.2× over v6.2, beats everything, pure local keywords

Search latency: ~7ms (keyword mode, measured on real data). Target: always < 50ms.

What's new in v6.4

• User profile reconstruction — multi-source distillation (decisions, tagged summaries, structured facts) instead of copying files
• True BM25 ranking (v6.3) + weight rebalance: retrieval is decoupled from memory importance
• 8/8 test suite passing · zero-dependency · zero API keys

Full details in CHANGELOG.md · install: cd Mnemosyne-v6.4 && bash install.sh

📊 How It Compares（full documents are in the /docs file）

💡 When to choose Mnemosyne: You want a memory system that costs nothing to run, keeps all data as human-readable files you can git diff, and works entirely offline without any external service dependencies.

⚠️ When NOT to choose Mnemosyne: You need cross-user shared memory at scale, or require deep semantic understanding beyond TF-IDF keyword matching.

If you want you get more information please step to MNEMOSYNE-REFERENCE.md

If you want to get more about the product iteration information please step to CHANGELOG.md