Every conversation with an LLM resets memory. Stash changes that by giving agents persistent memory across sessions. It plugs in as a cognitive layer between your chatbot and external data, turning raw user inputs into structured knowledge. Under the hood it uses Postgres with pgvector for storage, plus a background service that runs eight consolidation stages—fact extraction, relationship mapping, causal links, goal tracking, failure patterns, hypothesis checks and confidence decay. Each stage only processes new entries, so your knowledge base grows efficiently over time. Getting started takes three commands. Clone the repo, copy and edit .env.example with your API key and model choice, then launch everything with docker compose up. You get a fully self-hosted stack—Postgres, migrations, pgvector indexes and an MCP-compatible server. That means you can use Stash with any agent that speaks the Model-Control-Plane protocol: OpenAI Agents, Claude Desktop, Continue, Cursor, Windsurf, Ollama, OpenRouter or any other MCP client. The project is open source under Apache 2.0. You can tweak consolidation rules, swap models or integrate custom retrieval logic. By logging every interaction as “episodes,” then elevating episodes to facts and patterns, you avoid repeating context in each session. Over time, your agent actually learns your preferences, ongoing tasks and long-term goals.