Mems

A layered memory infrastructure for AI agents — remembers, recalls, distills, and archives.

Mems stores and recalls memories across four layers. The caller owns runtime state.

• L0 handles hot recent-memory cache
• L1 stores episodic memory — the source of truth
• 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

• ⚡ Hot memory cache (L0) with Redis-backed fast recall
• 🧠 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 Remember -> Recall