omk-context-broker

To maintain context for long-running multi-file coding sessions, you need a context broker policy that organizes project-local memory and run-local checkpoints across turns and workers. This prevents losing critical state across retries and multi-agent handoffs.

The best way to prevent losing state during risky refactors and multi-agent handoffs is to use Okabe or D-Mail style checkpoints before destructive actions. This triggers smart recovery and ensures durable memory file conventions for project facts and decisions.

Local graph memory works by acting as the default source of truth, reading and writing stable project facts and temporary run or session facts. It maintains a local mirror at .omk/memory/ to keep agent memory coherent across repeated project runs.

You need checkpointed recovery for multi-agent coordination when tasks span many files, turns, or DAG workers and face memory compaction moments. It ensures durable recovery before destructive actions and maintains state across complex project work.

Yes, omk-context-broker requires project-local graph memory tools for stable facts and run-local memory for temporary state. These dependencies are necessary to establish the local mirror at .omk/memory/ and execute Okabe or D-Mail checkpoints effectively.