dsh-knowledge-sync

A DeepSeek Harness conversation ends, and everything it produced—the questions asked, the tool chains run, the decisions made—vanishes the moment the session closes. Open the same project in the next round and the agent has no memory of any of it; you paste the context again or watch it re-diagnose a build failure you already fixed. dsh-knowledge-sync exists to end that start-from-zero-every-time cycle.

The approach is deliberately restrained. At the end of each round, a salience filter decides whether the exchange is worth keeping—only rounds with a substantive conclusion are frozen into a single Markdown file with YAML front matter. Before the file touches disk, tool arguments are scrubbed of tokens, password values, .env secrets, and PEM keys, with extra regex patterns configurable. Documents come in three granularities: raw captures the full question, answer, and tool trace; note is a finding the agent chooses to record mid-work; distilled condenses a long round into a summary via one small model call (off by default). Files live under ./knowledge in the workspace root—grep-able, git-ready, with no hidden sidecar index.

Recall is designed to cost almost nothing. The plugin does not paste the knowledge base into the system prompt, which would eat the context window and invalidate the prompt prefix on every write. Instead the agent sees a single pointer—N documents from earlier rounds exist in this workspace—and calls knowledge_search when the task looks familiar. The search runs through an in-memory BM25 index that covers full document bodies with CJK bigram support, not just titles, and returns a snippet per hit. Results can be filtered by kind, tag, or tool, with notes and distilled findings ranking above raw transcripts. Every module—capture, note, recall, distill, http—is independently optional, so you can run headless without a web page or skip distillation while keeping the rest.

For anyone who lives in the same project day after day, iterating and debugging, this conversation-as-document pattern turns repeated context into a small, searchable library the agent consults only when it genuinely needs to—no injection, no bloat, just a quiet file on disk that says the answer was already found.