dsh-plugin-memos-code-retrospect

AI coding assistants have a maddening habit of falling into the same trap twice, resurfacing a solution that was just rejected under slightly different wording, and forgetting the project architecture conventions the moment a new session starts. dsh-plugin-memos-code-retrospect was built to close that memory gap.

The design is deliberately lean. At the end of every turn it calls an LLM to distil three kinds of entities from the full conversation context - rejected solutions, debugging pitfalls, and project-level engineering constraints - then applies value filtering and similarity-based deduplication before writing them into the MemOS memory substrate. Right before the next agent inference step, if the task is classified as coding, memories tagged rejected_solution are retrieved with a boosted weight and stitched into the prompt so the model actually remembers why a previous approach failed. Non-coding tasks skip the injection entirely to keep other workflows free of coding-memory noise.

This plugin is aimed at developers who drive coding workflows through DSH (DeepSeek Harness / Cordis) and are tired of an AI that keeps making the same mistake. It also doubles as a ready-made ablation switch for research: flip the enabled flag to compare runs with and without the retrospective loop. All persistence, vector search, archival, and eviction are delegated to MemOS; the plugin itself stores nothing, keeping deployment light and the path to production straightforward.