dsh-memory-graph

A DeepSeek Harness agent loses context the moment a session ends: user preferences, hard-won facts, constraints, and lessons learned simply vanish. dsh-memory-graph fills that gap with a local-first, durable memory layer backed by SQLite and a temporal knowledge graph. Stable facts are superseded rather than overwritten, every record carries full provenance, and any entry can be reverted to its previous version with a single call. Retrieval blends lexical match, graph proximity, importance, time decay, and explicit access reinforcement into one bounded ranker, and a dedicated FTS5 trigram index plus short-query fallback handles continuous CJK text without whitespace tokenization. An optional OpenAI-compatible embedding endpoint can layer cosine similarity on top, while the system degrades gracefully back to lexical and graph recall when no key or endpoint is available. The plugin plugs into DSH purely through lifecycle extension points—no source patch, no agent-loop fork—and unloading it strips every tool, listener, route, and UI contribution in one step. On the web side, an interactive graph view supports search, drag, zoom, entity merge/rename/delete, orphan discovery, and per-turn Memory On / Off / Mixed latches so you can dial automatic recall, summarization, and archiving on or off for the very next turn without touching global defaults. It is a natural fit for developers and teams running long-lived multi-session agents, those who keep all data on their own host, and anyone whose conversations are predominantly in Chinese.