bmem

What bmem is

bmem is an open catalog of web skills — reusable, battle-tested recipes for operating specific sites (search LinkedIn, read a Reddit thread, …). A skill ships the exact selectors and, when possible, the direct in-page XHR, so you don't have to read a whole DOM to figure the site out. You run the skill with your own browser; bmem never drives a browser and never sees your cookies.

Use this loop whenever a task involves operating a website.

The loop

1. Search first. Before exploring a site by hand, look for a skill:

   bmem search <query>              # e.g. bmem search linkedin people
   bmem search --site linkedin.com  # filter by site
   bmem list --limit 50             # browse the whole catalog
   

   Results are tiny index entries: name, intent, side_effect, needs, verified.

   Want the full recipe before committing? Read it without installing:

   bmem show linkedin.com/search-people   # prints the SKILL.md, no install
   

2. Add the skill. If one fits, register it as a native skill:

   bmem add linkedin.com/search-people   # one skill
   bmem add linkedin                     # a bare site adds every skill it has
   

   After this the skill appears in your own skill list — invoke it natively. You do NOT call bmem again to read it.

   Never invent a skill id. Only add a name that came back verbatim from search or list. The catalog is curated and read-only — a made-up id like linkedin.com/find-jobs will just 404. If nothing matches, go to step 4.

3. Run it with your browser (see Execution below).

4. No skill found? bmem search is empty → explore the site yourself, find the real XHR endpoint in the network log (more robust than clicking), and author a new SKILL.md. Authoring is out-of-band — see the project's authoring guide.

Reading a SKILL.md

A skill declares the browser capabilities it needs (needs: [navigate, evaluate]) and writes its ## Execute steps in those generic verbs. It never names a concrete tool — you map each capability to whatever browser you have.

• needs: [evaluate] (± navigate) — the robust case: one in-page fetch + parse.
• needs includes click/type/snapshot — UI-driven, more fragile. Requires a browser that can drive the rendered page.

Check needs up front: if your browser can't provide a capability, switch to one that can.

Capability → tool mapping

Translate each capability verb to your browser's tool:

Skills ship a CSS selector / description, never an ephemeral ref or coordinate — resolve it at runtime. Prefer doing DOM actions inside evaluate (document.querySelector(sel).click(), scrollIntoView()) — it's uniform across every browser and beats native click/type.

Execution

One path, browser-agnostic:

1. navigate to the URL in the skill's Execute step 1 (establishes the session / same-origin for the fetch).
2. evaluate the skill's ## Extractor function with the skill's params. It does the in-page authenticated fetches and parsing, and returns the JSON described in returns.
3. Read the result back via your browser's evaluate-result channel and check the skill's ## Success assertion.

The browser holds the session — no cookies ever leave it.

Auth (auth: session)

Some skills need a logged-in session. Check the skill's ## Preconditions:

• If the site is already logged in in the browser profile you drive → just run it.
• If not → navigate to the site, let the user log in manually, then retry. The session persists in the browser profile; don't try to automate the login.
• Expired session mid-run → re-prompt the user to log in, then retry. Don't re-author.

Self-host / staging

Point the CLI at another catalog with BMEM_SKILLS_API_BASE_URL. No config command needed.