dsh-experience-library

Memory plugins let your AI remember facts, yet it still stumbles over the same pitfalls in unfamiliar territory. dsh-experience-library solves a deeper problem: accumulating procedural knowledge. It captures zero-token signals from session events, tool failures, and retry traces, filters out hallucinations through a three-layer verification gate, and solidifies survivors into reusable skill books the next time a similar task appears.

Core mechanics include idle-window auto-dispatch for batch refinement, a locate index that eliminates redundant file lookups, an adapter layer that plugs any memory plugin's lesson store into a unified pipeline, and a skill-book library that grows organically from your own usage. The ten books shipped in the repo are reproducible samples; your library will accumulate its own expertise over time without manual curation.

Benchmarks show concrete impact: on complex session-log repair tasks, the experience library lifted success from the bare baseline of 60 percent to 100 percent, cut wall-clock time by 4.7 times, and reduced thinking tokens by 77 percent. The gain scales with task unfamiliarity, so it shines when the model faces unknown-domain pitfalls. This plugin suits DeepSeek Harness developers who repeatedly face high-error tasks and want their agent to stop re-learning what it already knows.