dsh-engram

Every new session with an LLM Agent starts from amnesia: the code style, naming conventions, and architectural preferences you described last week are simply gone. dsh-engram gives DeepSeek Harness cross-session long-term memory so agents can retain user preferences, project conventions, and factual history across conversations, evolving them over time through ingestion, reinforcement, distillation, and decay.

Under the hood it runs a two-tier store: user-level facts in a global database and project-level conventions keyed by repository origin, keeping personal and project memories cleanly separated. Retrieval blends FTS5 Chinese-aware keyword search with local bge vector embeddings via RRF fusion, augmented by relation-edge expansion and recency boosting; the embedding model downloads once and then operates fully offline, gracefully degrading to pure keyword search on failure. Every memory carries its source session, turn, and full audit trail; engram_review walks supersession chains and contradiction candidates. A built-in Web panel in Settings offers multi-criteria filtering, inline editing via the supersession chain, soft-delete recovery, and one-click Markdown or JSON export.

Built for developers who work with the same agent over many sessions and would rather not re-brief it on every architectural decision or coding convention. Pure TypeScript, zero external processes, zero Python dependencies, and fully portable data make it straightforward to adopt and easy to leave.