Granular context memory for DeepSeek Harness
vilicvane/dsh-plugin-turn-memory · skills
Turn-granular context memory for DeepSeek Harness.
★ 2GitHub stars
0Forks
2026-08-16Last updated
TypeScriptLanguage
—License
Key features
- older turns use a subagent, where a cheap model reads the turn's full text and answers the targeted question.
- Hindsight is written in natural language, not bracket tags: when a later development proves an earlier entry wrong, the correction is annotated inline at the point it went wrong — "I thought X might work.
- If the main agent leaves a turn unsummarized for a whole turn, a fallback fork summarizes it (one turn late)
- dsh-web-restart — restarting the dsh web server after profile/plugin changes.
- raw — the turn's full text is returned directly into the conversation
Install command
dsh plugin --profile web add github:vilicvane/dsh-plugin-turn-memory