Stash - A Persistent Memory Layer for AI Agents

An open-source project built around a persistent memory layer that carries conversation and task context across sessions. It stores raw observations as episodes and continuously synthesizes them into structured knowledge.

The core storage structure is based on PostgreSQL + pgvector, and it's designed to be model-agnostic so it can be attached to Claude, GPT, local LLMs, custom agents, and the like. Unlike document-retrieval-focused RAG, it differentiates itself by accumulating newly emerging facts, relationships, goals, failures, and hypotheses from conversations.

Memory is separated by namespace and hierarchical paths. For example, it's designed so that self-knowledge under users, projects, and /self doesn't get mixed together — reads include subpaths, but writes are recorded only to the exact namespace to prevent contamination.

A background consolidation pipeline is also provided.

• episodes: stores observations in an append-only manner
• facts: structures bundles of episodes
• relationships / causal links: extracts entity relationships and cause-effect links
• patterns: derives higher-level abstractions
• contradictions: contradiction detection and confidence decay
• goal inference / failure patterns / hypothesis scan: goal tracking, learning from repeated failures, hypothesis verification

It also supports MCP-native integration, allowing it to connect to multiple clients such as Claude Desktop, Cursor, and OpenCode. It provides a total of 28 tools, ranging from basic functions like remember, recall, forget, and init to functions related to causal links, contradictions, and hypotheses.

An agent uses init to create a /self namespace, then builds a self-model by accumulating capabilities, limits, and preferences. The project also presents a flow with a 5-minute-interval research loop that pulls in past context and automatically repeats web research, hypothesis generation, consolidation, and summarization.

Deployment is provided via Docker Compose, and the project is released under the Apache 2.0 license.