ABOUT

Long-term memory for your AI

Rohrpost is an open-source knowledge base that gives AI assistants persistent, searchable memory — deployed on your own infrastructure.

Why Rohrpost?

LLM context windows are finite. Every conversation starts from scratch, and every session ends with the model forgetting everything it just learned about your codebase, your architecture, your decisions. The longer a project runs, the worse this gets.

Rohrpost solves this by continuously capturing what matters — decisions, entities, code references, architectural context — into compact capsules. When you start a new conversation, the most relevant capsules are recalled and packed into a token-efficient brief that your AI assistant reads before answering. The result is an assistant that remembers what you told it three weeks ago.

Design principles

  • Your data stays yours. Everything runs on-prem. No telemetry, no call-home, no licensing servers. The embedding model (Ollama) runs locally by default.
  • No custom integration required. Rohrpost speaks MCP — the open protocol Claude, Cursor, VS Code Copilot, and other AI tools already support. Configure it once, and any compatible assistant can read and write your knowledge base.
  • Hybrid retrieval over pure vector search. Flat vector similarity misses too much. Rohrpost combines vector search (pgvector), keyword matching (BM25 via Tantivy), and 1-hop graph expansion (Apache AGE), merged with Reciprocal Rank Fusion and cross-encoder reranking.
  • Operationally simple. A single ./deploy/deploy.sh up starts the entire stack — Postgres, NATS, Ollama, Garage, and all Rohrpost services. No Kubernetes required to get started.

The stack

Rohrpost is built with boring, reliable technology:

  • Rust — all services compile to a single ~30 MB static binary
  • Postgres + pgvector — chunk storage and vector similarity search
  • Apache AGE — property graph database for entity relationships
  • Tantivy — embedded BM25 full-text search
  • NATS JetStream — durable message queue with at-least-once delivery
  • Garage — S3-compatible object store for raw capsule archives
  • Ollama — local embedding model (no external API required)
  • axum — async HTTP server for the gateway and admin API
  • rmcp — Model Context Protocol server

Open source

Rohrpost is developed in the open on Codeberg. Contributions, bug reports, and feature requests are welcome.

View on Codeberg →

Get in touch

Found a bug? Have a question about self-hosting? Want to share how you're using Rohrpost? Open an issue on Codeberg or reach out on the contact page.