dsh-learn-wiki

Most agent memory plugins are a static pipeline: if retrieval misses, it just misses, and the same knowledge gap gets re-hit every single session. dsh-learn-wiki adds one crucial increment treat repeated struggle as an observable signal, trigger background rate-limited web acquisition, distill the result into a staging area, promote it into a local knowledge base through a human-gated two-stage commit, and auto-recall it before the next work session. It ports the Corrective RAG feedback loop from single-turn Q&A onto the agent loop: if the wiki does not know, go learn; once learned, it stays.

The core loop spans three tiers: L1 is a standalone Markdown repo of curated knowledge pages injected directly into the prompt on a strong hit; L2 is handled by the existing hindsight plugin for episodic memory recall, which this plugin deliberately does not reimplement; L3 is the web fallback that rate-limits fetches through the host web interface and extracts article bodies. The default trigger is struggle (consecutive failures, hitting the same wall, fixing but still failing), not the cheaper retrieval miss, because any novel topic will miss and would cause the system to fetch for every new thing mentioned. Staged content never participates in auto-recall, and every entry must be traceable to a URL or file, keeping the contamination surface confined to the staging area.

A set of explicit tools (wiki_recall, wiki_harvest, wiki_sessions, wiki_commit, and more) complete the picture. wiki_sessions reads local historical sessions to surface recurring failure patterns and a hand-off brief, turning the question of how I solved this last time from unanswerable into a one-liner. A sidebar panel with three tabs (capabilities, knowledge, acquisition) plus offline-runnable UI self-test scripts mean nothing requires a browser or a host restart. Built for developers already running multi-turn agent workflows in DSH who want each knowledge gap hit exactly once.