Introducing Rohrpost: Context Compression for LLM Workflows

If you’ve spent time building with large language models, you’ve hit the wall: conversations grow, context windows fill up, and suddenly your AI assistant has forgotten everything you discussed twenty minutes ago.

The problem

Every LLM interaction carries a token budget. Once you exceed it, you’re forced to choose between:

• Truncating history — losing earlier decisions and context
• Summarizing manually — time-consuming and error-prone
• Starting fresh — repeating yourself from scratch

None of these options are great when you’re deep in a multi-day coding session or managing a complex project.

Our approach

Rohrpost takes a different path. Instead of fighting context limits, we work with them. The system continuously distills your conversation history into compact, structured capsules that preserve:

• Architectural decisions and their rationale
• Code references and file relationships
• Project status and open questions
• Key entities and their connections

When you need context for your next LLM call, Rohrpost recalls only the capsules relevant to that query and assembles a compact brief — a fraction of replaying the full conversation history, because it’s targeted instead of exhaustive.

How it works

Raw conversation
    ↓ distill (chunk, embed, extract entities)
Capsules + chunks + graph edges
    ↓ recall (vector + BM25 + graph, ranked and packed)
Brief for next prompt

The compression isn’t just truncation. Rohrpost uses a hybrid retrieval pipeline — vector similarity, BM25 keyword search, and graph traversal — to identify what actually matters for the question you’re asking right now, not everything that was ever said.

What’s next

We’re actively developing Rohrpost in the open. Check the changelog for the latest releases, or dive into the documentation to try it yourself.