🧠 Engram

Cognitive Memory System for AI Agents

Persistent semantic search · Knowledge graph · Sleep consolidation · Zero external APIs

Quick Start · Architecture · Features · Research · CLI · Agent Integration

> **Engram** gives AI agents a brain — not a database. > It remembers what matters, forgets what doesn't, and builds connections between ideas. > All inference runs **locally** — no API keys, no cloud, no token cost per memory operation.
## Why Engram? Most AI memory solutions (like [Mem0](https://github.com/mem0ai/mem0)) use LLM calls for every `add()` — extracting facts, classifying operations, resolving conflicts. That's powerful, but **expensive** and **opaque**. Engram takes a different approach: | | **Mem0** | **Engram** | | ---------------------------- | ------------------------------ | ----------------------------------------- | | **Who decides what to save** | LLM automatically | Agent explicitly | | **Cost per memory write** | LLM call (extraction + update) | Local embedding only | | **Cost per memory read** | Vector search + LLM | Hybrid search + local reranker | | **External dependencies** | OpenAI API / vector DB service | **None** — fully local | | **Memory unit** | Atomic fact ("User prefers X") | Typed record with full context | | **Forgetting** | No built-in mechanism | Ebbinghaus-inspired decay + consolidation | | **Knowledge graph

Agents

engram-ai-memory

vestige