Stash: Open source persistent memory layer for AI agents

Stash is an open-source, Apache 2.0-licensed persistent memory layer for AI agents, designed to give them long-term memory across sessions. Unlike retrieval-augmented generation (RAG), which only searches static documents, Stash learns from every conversation, decision, and failure, synthesizing raw observations into facts, building a knowledge graph, detecting contradictions, and tracking goals. It uses a namespace system for organizing learned information hierarchically, similar to folders on a computer. Stash works with any MCP-compatible agent (e.g., Claude Desktop, Cursor) and any OpenAI-compatible backend (cloud, local, or self-hosted via Ollama, vLLM, LM Studio, Together AI, Groq). It provides 28 tools covering the full cognitive stack, from raw recall to causal chains and hypothesis management. A background process continuously synthesizes experiences into structured knowledge. Stash is backed by PostgreSQL and pgvector, with Docker Compose handling setup. It supports multiple models via OpenRouter or local LLMs like Qwen, Llama, Mistral. The project is hosted on GitHub and aims to make memory a universal feature, not a premium lock-in.