Mems

A layered, self-distilling, century-scale memory system for AI agents.

Mems is an industrial-grade memory hub solution for AI agents. Through a four-layer hot/cold-decoupled architecture, it provides a memory foundation with personality consistency, low-cost retrieval, and structured evolution.

• L0 handles live session working memory
• L1 stores episodic history and active online recall data
• L2 distills reusable long-term knowledge
• L3 archives old memory into JSONL

Start Here

• Quick Start
• Integration Overview
• Architecture Overview
• Benchmarks

Core Design Philosophy: Why Mems

• ⚡ Instant working memory (L0) with Redis-backed live context
• 🧠 Asynchronous knowledge distillation (L2) from raw L1 history to stable semantic facts
• 🔍 Evidence grounding so distilled memory remains traceable to L1 source material
• 📜 Century-scale archive (L3) with text-first JSONL retention
• 🛡️ Production-grade isolation across tenant, user, agent, and session boundaries
• 🔌 Atomic integration through Context -> Query -> Write