llm‑wiki improves agent recall

Most artificial intelligence agents do not really remember; they re-search. A retrieval system called retrieval-augmented generation pulls text snippets at question time, so the model often has to rebuild the answer from scratch on every turn. (github.com) The “wiki” idea flips that. Instead of fetching raw fragments every time, the model writes a persistent set of linked markdown pages, updates them when new sources arrive, and queries that compiled knowledge later. (github.com) That is the pattern behind llm-wiki, an open-source project by nvk that ships as a Claude Code plugin or as an AGENTS.md file for other coding agents. Its README says it handles source ingestion, wiki compilation, querying, research, and output generation from the same knowledge base. (github.com) The weak point in systems like this is routing. If a user types “where was I,” “research export controls,” or pastes a URL, the agent has to decide whether that input means resume, research, ingest, or query before it can touch the right memory. (github.com) Version 0.0.19 changed that piece. llm-wiki added what it calls a “Fuzzy Intent Router,” so the `/wiki` command can understand natural-language requests and map them to the right subcommand instead of forcing the user to remember exact syntax. (github.com) The project’s examples are concrete: a pasted URL routes to ingest, a question routes to query, “research X” routes to research, and “where was I” routes to resume. When the input is unclear, the tool now shows a numbered menu instead of guessing wrong. (github.com) That sounds small, but brittle agents often fail on exactly this kind of front door problem. A memory system can be full of useful notes, logs, and summaries, yet still feel forgetful if the user’s wording misses the one command shape the tool expects. (llm-wiki.net) llm-wiki had already added a resume feature in version 0.0.18 that reloads interrupted sessions by reading recent log entries, wiki statistics, and last-updated articles. The new fuzzy router makes that memory easier to reach because a user can ask in plain English instead of recalling a flag like `--resume`. (llm-wiki.net) There was one catch right after the release: version 0.0.20 fixed the bare `/wiki` command so the fuzzy router from 0.0.19 was actually reachable without typing `/wiki:wiki`. In other words, the idea landed in 0.0.19, and the next release fixed the doorway into it. (github.com) This is why the update matters for agent design. Better memory is not only about storing more facts; it is also about matching messy human requests to the right stored context, so the agent can find the page, log, or research thread you meant instead of acting like it forgot. (github.com (llm-wiki.net))