Granular context memory for DeepSeek Harness

Project overview

Turn-granular context memory for DeepSeek Harness: step 1 of a two-step context-compression plan.

The default mode is agentic; the time routing is internal to the tool.

A conditional tail reminder backs it up: once the current turn spans more than reminderNodeThreshold surface nodes (default 30, counted by nodes, not tokens), the end-of-context runtime snapshot carries one stable line nudging the model toward compactturn; at 1.5x the threshold the line escalates to a direct warning. Below the threshold the reminder contributes nothing — zero tokens, zero noise — and it disappears on its own once a compaction lands. Summary format (version 4)