1. Every session starts from zero. The agent sees your code but has no idea why it looks the way it does.                                                      
  2. Team knowledge stays siloed. What one developer's agent learned yesterday doesn't help another's today.                                                     
                                                                                                                                                                 
  Rekal captures session context (turns, tool calls, file changes) at every commit into a local DuckDB database. A 2-10 MB session file compresses down to ~300  
  bytes on the wire via a custom binary codec with zstd and string interning. It shares across your team via git orphan branches — no extra infrastructure.      

  Search is three-way hybrid (BM25 + LSA + nomic-embed-text), all inside a single binary. No external APIs, no cloud, no setup. The embedding model ships in the
  binary.

  The agent controls how much context it loads — progressive retrieval keeps token cost minimal.

  Other design choices: append-only with content-hash dedup (merge conflicts are structurally impossible), ~200ms search at 14k turns, zero runtime dependencies
  beyond git. Written in Go, Apache-2.0.