myc run linkedin.com
Public-page fallbacks
Remember when search results beat deep navigation.
// agent memory kit
Give browser agents a second run.
Mycelium wraps your web agent, records what happened, extracts reusable domain hints, and primes the next attempt before the browser opens.
agent@run:~$ myc run linkedin.com "summarize post"
[prime] no knowledge found - starting fresh
[record] saved hint: search result before deep nav
agent@run:~$ myc run linkedin.com "summarize post"
[prime] 1 hint loaded before the agent starts// why use mycelium
Agents are good at browsers. Bad at remembering browsers.
Most web agents start cold on every run. They retry the same login wall, rediscover the same public fallback, and spend tokens learning routes that were already found yesterday.
Mycelium turns completed runs into a local memory graph: domains, outcomes, symptoms, routes, and reusable hints that can be retrieved before the next attempt.
// live demo run
Watch a cold run become a primed run.
The first run starts empty, finds a working route, and records the useful part. The second run loads that hint before the agent spends a token in the browser.
01
Cold start
No memory found for the domain.
02
Record
Search fallback worked; save the route.
03
Prime
Load the hint before the next run.
$ myc run linkedin.com "summarize a post"
- no knowledge found - starting fresh
+ saved hint: try search results before deep navigation
$ myc run linkedin.com "summarize a post"
+ 1 hint loaded before the agent starts
// run lifecycle
Prime, run, record.
Mount it around Stagehand, Browser Use, TinyFish, or your own browser loop. The agent still gets a plain task; Mycelium handles the memory before and after.
01Primeload domain hints from the graph
02Runhand the enriched task to an adapter
03Recordstore outcomes, symptoms, and routes
04Retrievesemantic match for the next similar goal
// use cases
The memory agents wish they had.
myc inspect checkout.app
Blocked-flow symptoms
Stop replaying paths already marked as login-gated or brittle.
myc history docs.site
Domain memory
Keep successful routes close to the project instead of in a prompt doc.
myc stats
Run intelligence
See hints loaded, hints extracted, and which domains are improving.
npx myc-explorer
Memory dashboard
Open a local UI for domains, runs, hints, patterns, and prompt previews.
adapter: stagehand
Adapter-agnostic runs
Use the same prime and record contract across browser tools.
store: sqlite
Local-first retrieval
Semantic search works from the project graph without a hosted database.
// explorer dashboard
Inspect what your agents remember.
Mycelium Explorer is a local dashboard for the same SQLite graph used by the SDK and CLI. Browse learned domains, compare runs, review extracted hints, inspect patterns, preview primed prompts, and open benchmark results without shipping memory to a hosted service.
// what changes
The second attempt starts with context.
No domain memory found. Agent explores, fails, retries, and records the result.
Failure patterns and workarounds become reusable facts in the graph.
Relevant hints are injected before the next browser session starts.
// spec sheet
Small pieces. Plain contracts.
// sdk
Drop it around an agent run.
import { run, stagehandAdapter } from "@mycelium-dev/mycelium"
const result = await run({
url: "linkedin.com",
goal: "open a public post and summarize it",
adapter: stagehandAdapter(),
})
console.log(result.primed.hintsLoaded)
console.log(result.recorded.hintsExtracted)No hosted vector DB required.
Runs, hints, outcomes, and relationships live in a local SQLite graph.
LLM paths are optional.
Use deterministic classification by default, then opt into OpenAI extraction when useful.
CLI and dashboard included.
Use myc for terminal inspection or myc-explorer to browse memory visually.
// install