dsh-llmwiki

Every DeepSeek Harness conversation is bounded by the context window; the moment a session ends, the judgments, preferences, and domain notes accumulated with the model simply vanish. dsh-llmwiki treats a local Markdown vault as a disk-level memory store, giving the model a persistent, cross-session memory layer that activates automatically on every turn.

Retrieval goes beyond plain-text matching. Three strategies- keyword search, an Obsidian-style wikilink graph, and a temporal look-back window- run in parallel, are fused via Reciprocal Rank Fusion, and assembled within a configurable token budget before injection. The model can also proactively recall or persist insights through memory_search and memory_save tools, while every turn is auto-appended to a date-stamped chronicle file, building a searchable history. The entire plugin ships with zero runtime dependencies and works out of the box, with knobs for strategy selection, time windows, cache TTL, and assembly priority when finer control is needed.

If you already manage personal knowledge in Markdown or Obsidian and want your AI assistant to genuinely remember, dsh-llmwiki bridges the two: your notes stay local, the model calls on them, and you just keep writing.