Stash — Self-Hosted Memory + MCP SSE on Postgres/pgvector

Practical Notes

• Quant: the README highlights an 8-stage consolidation pipeline—treat that as a controllable knob for memory quality vs cost.
• Quant: use one SSE endpoint per environment (dev/staging/prod) so memory doesn’t cross-contaminate across contexts.

Rollout pattern

• Start with one agent and one project; define what counts as “write-worthy” memory (facts, decisions, constraints).
• Add retention and consolidation schedules once you observe retrieval precision in real tasks.
• Create a “memory schema” (tags, owners, environments) so retrieval stays scoped and predictable.

Watchouts

Memory is also a data-leak vector. Treat the store as sensitive: enforce auth, avoid storing secrets, and keep environments separated.

FAQ

Q: Is this only for Cursor? A: No. It works with any MCP client that supports SSE endpoints; the README lists multiple clients.

Q: How do I keep memory useful? A: Write only verified facts/decisions and enforce tags + scope per agent or per project.

Q: Can I run it in production? A: Yes, but add auth, backups, and monitoring; don’t leave it anonymously accessible on a network.