Stash is an open-source, self-hosted MCP memory server backed by PostgreSQL + pgvector that gives any AI agent persistent, structured memory across sessions.
Key Takeaways
Ships 28 MCP tools covering the full cognitive stack: raw remember/recall up through causal links, contradiction resolution, goal inference, and hypothesis management.
Memory is organized via hierarchical namespaces (e.g. /users/alice, /projects/restaurant-saas, /self/limits); reads from a parent path recursively include all child namespaces.
The 6-stage background consolidation pipeline promotes raw episodes into facts, relationships, causal links, patterns, and contradictions without manual intervention.
Supports any OpenAI-compatible inference endpoint: OpenRouter, Ollama, vLLM, LM Studio, Groq – same provider config serves both the embedding model and the reasoning model.
Critical footgun: STASH_VECTOR_DIM must match your embedding model before first run; pgvector locks the dimension at initialization and changing it later requires a full database reset.
Hacker News Comment Review
Only the author commented at time of fetch, reiterating that platform memory (Claude.ai, ChatGPT) is model-locked while Stash targets any agent via MCP with Apache 2.0 licensing.
No substantive independent discussion of tradeoffs, failure modes, or production experience yet – the consolidation pipeline’s LLM cost at scale and the irreversibility of STASH_VECTOR_DIM are the most obvious open questions not addressed in comments.
