Engram

A three-layer memory system that mimics how humans actually remember things. New memories hit a buffer, get promoted to working storage if they matter, and eventually land in core if an LLM decides they're worth keeping long term. Memories decay at different rates (episodic fade fastest, procedural slowest) and get reinforced when recalled. It auto-merges duplicates, clusters related memories into topic trees, and lets you tag memories as triggers so your agent can pull deployment lessons before actually deploying. The /resume endpoint gives you full context on startup, /recall does semantic search, and /topic lets you drill into specific clusters. Single Rust binary backed by SQLite, no vector database required.