dsh-mnemosyne-memory

Every fresh session means your AI Agent starts with amnesia—architectural decisions, hard-won lessons, and the codebase layout from the previous conversation all evaporate, forcing you to re-brief it by hand. Mnemosyne closes that gap for DeepSeek Harness by persisting session knowledge and enabling vector-based semantic recall, so the Agent genuinely remembers across sessions and restarts with zero manual intervention.

Under the hood, the plugin ships 11 tools spanning the full memory lifecycle: event storage and semantic retrieval, LLM-driven deep reflection that extracts decisions and insights, automatic knowledge-page generation with incremental delta refresh, and a codebase survey that indexes 30-plus configuration patterns. It supports per-workspace isolation or team-shared memory, and runs on a built-in cadence—reflecting every 5 turns, refreshing pages every 10 turns, and injecting relevant history before each step—so the Agent grows its understanding of your project over time.

If you run multi-turn development with DSH, want your Agent to accumulate project knowledge over time, and would rather not pay for embedding APIs, Mnemosyne is the zero-cost answer. Go fully offline with local Ollama, or tap Google AI Studio's free tier; both paths cost nothing. Privacy-sensitive environments, offline setups, and teams that need to share project context are all natural fits.