Knowledge & Open Capabilities

Knowledge bases

A knowledge base is NomiFun’s “config one, use anywhere” idea applied to content: a markdown directory you curate, mounted into a session workspace so the agent retrieves before it answers per an explicit protocol, rather than answering from memory. A base can bind to a session, a terminal, a specific companion, or a working directory, and can be exported / imported as a .zip.

Entry point: the Knowledge item in the left sidebar (route /knowledge). Everything below — creating bases, setting write-back, binding, and mounting — lives here.

Knowledge base home

Three knowledge sources

  1. Local markdown — a directory you organize yourself; the most direct option, ideal for handbooks, specs, and project notes.
  2. Live URL snapshots — supply up to 16 public URLs at creation. Snapshot mode fetches each page now and converts it to markdown into the base (oversized pages compressed by AI); Live mode leaves fetching to the agent at runtime. Built-in SSRF protection accepts only http/https, and JS-heavy pages are fetched through a headless browser to recover their content.
  3. Feishu — connector implemented (the in-UI creation entry is currently off).

Notion source = roadmap, not yet implemented. The available sources today are local markdown, URL snapshots, and the Feishu connector.

Steps

  1. Create a local base. On the knowledge page, click new, give it a name, organize or select the markdown directory, and optionally add URL sources.

    Create a local knowledge base

  2. Add live-snapshot URL sources (optional). Paste up to 16 public links and choose Snapshot or Live per link. Snapshot suits stable reference pages; Live suits pages where you want the latest version every time.

    URL live-snapshot knowledge base

  3. AI-generate a description. Click “AI generate” to call autogen and produce the base’s description and README.md (requires a configured AI provider). The description feeds the agent’s “when to consult” hint — the clearer it reads, the better the matches.

  4. Configure write-back mode and settings. In base settings, decide how knowledge produced in conversation flows back:

    • Disabled — no write-back.
    • Staged — write-back lands in the base’s _inbox/ for you to review before merging. Writes from IM always go Staged, never direct.
    • Direct — skips staging and writes straight into the base body.

    Knowledge base settings (write-back mode, etc.)

  5. Inspect the base detail. After creation, open the detail page to browse the directory tree, fetched snapshot files, pending _inbox/ entries, and the base’s description and stats.

    Knowledge base detail

  6. Bind / mount it to a use case. Bind the base to the current session, a terminal, a companion, or a working directory. Once bound to a companion, its companion chats and channel sessions all mount that companion’s bases automatically; once mounted into a session workspace, the agent retrieves-before-answering inside that session.

    Mount into a session workspace

Four binding types

Binding typeMeaning
workpathBound to a working directory; sessions under it mount by default
conversationBound to this one session only
terminalBound to a terminal session
companionBound to a companion; its companion chats and channel sessions mount it throughout

What the agent sees

Bases mount at {workspace}/.nomi/knowledge/. The injected context carries each base’s description, an AI summary, a “when to consult” hint, and a budgeted table of contents (20 entries per base / 60 global), plus an explicit retrieval protocol that steers the agent to retrieve before it answers. Companions can also grow their own bases — the Desktop Gateway provides knowledge tools (create base / write file / fetch URL), so a companion can quietly distill notes during a chat.

Notes & boundaries

  • Any agent can use it: once a base is bound to a session / terminal / working directory / companion, whatever agent runs in that context — the built-in nomi, an ACP direct-connect agent (Claude Code / Codex, etc.), or a terminal CLI — retrieves-before-answering by the same protocol. It is not tied to one kind of agent.
  • Safe write-back is a hard rule: any write originating from an IM channel lands in _inbox/ for review first and never goes straight to the body; you decide whether to merge it from the detail page.
  • URL fetching is SSRF-guarded: only http/https is accepted, and internal/loopback addresses are blocked.
  • Retrieval has a budget: the injected table of contents caps at 20 entries per base and 60 global; anything beyond that doesn’t enter context — organize the most important material up front.
  • Source status: local markdown, URL snapshots, and the Feishu connector are available; Notion is roadmap.

FAQ

Q: How do I choose between Snapshot and Live? A: Use Snapshot for reference pages that rarely change (fetched once, stable and predictable); use Live for pages where you need the latest version each time (fetched at runtime, higher cost).

Q: Where does the note I asked a companion to record over IM end up? A: In that base’s _inbox/, waiting for you to review and merge from the detail page before it reaches the body. This safeguard cannot be bypassed.

Q: Can one base serve multiple sessions at once? A: Yes. Bind by workpath or companion to let multiple sessions share the same base.

Full docs → GitHub