comparison

browser-memory vs Stagehand

Both help an AI agent act on real websites without brittle selectors. Stagehand asks an LLM to resolve each action live, every run. browser-memory remembers the resolved action as a deterministic, shared skill, so it never pays to re-derive it.

browser-memoryUpdated Jul 2026
browser-memoryStagehand
Core ideaShared memory of pre-built skillsAI SDK that resolves actions live
Per-run behaviorCalls a saved, deterministic toolAn LLM re-resolves the action each run
Token cost~4x fewer, no inference to actLLM inference on every act()/extract()
DeterminismSame tool, same resultNon-deterministic LLM resolution
When the site changesRepaired once, for every agentLLM adapts live, per agent, per run
Reuse across agentsShared catalog, install per taskYour own codebase
Output to the modelStructured data (fields, rows)Structured data via schema
Installbmem add {site}/{task}npm i @browserbasehq/stagehand

The real difference: resolve once vs resolve every run

Stagehand is a genuinely good AI automation SDK. It drops hardcoded selectors by letting an LLM interpret “click the submit button” at runtime, so scripts survive redesigns. But that resolution runs on every single call, which means model latency and tokens each time the agent acts on a page it has already seen.

browser-memory resolves the action once and saves it as a deterministic tool in a shared catalog: tested, repaired for everyone, and called with no inference to decide what to do. The agent gets fields back, fast and cheap, and the same tool works for every agent that installs it.

When to use which

Choose Stagehand for novel or one-off flows where you want an LLM to work out the page live. Choose browser-memory for the tasks you run repeatedly on real sites, where re-deriving with an LLM every run is wasted — about 20x faster and 4x fewer tokens once the action is a saved skill.

FAQ

What is the difference between browser-memory and Stagehand?

Stagehand is an open-source AI SDK (act, extract, observe, agent) that uses an LLM to resolve natural-language browser instructions at runtime, so it survives redesigns without hardcoded selectors. browser-memory is a shared memory of pre-built, deterministic skills reused over MCP. Stagehand decides the action live on every run; browser-memory remembers the resolved action as a saved tool so it is never re-derived.

Which uses fewer tokens, browser-memory or Stagehand?

browser-memory, on repeated tasks. Stagehand runs LLM inference every time it resolves an act() or extract(), so tokens scale with every run. browser-memory calls a saved tool that returns structured data with no inference to decide the action, which is where the roughly 4x token reduction comes from.

Is browser-memory a replacement for Stagehand?

They can be combined. Stagehand is great for novel or one-off flows where you want an LLM to figure out the page live. browser-memory is for the actions you run repeatedly on real sites, where re-deriving with an LLM every run is wasted latency and tokens — once resolved, the action becomes a shared, self-repairing skill.

Does browser-memory work with Claude, MCP and other agents?

Yes. browser-memory exposes skills over the Model Context Protocol, so Claude Code and any other MCP-capable browsing agent can discover and run them directly.

Compare more: browser-use, Playwright, Browserbase, Unbrowse.

Stop re-deriving the same page

Give your agent skills that are resolved once and reused forever, 20x faster with 4x fewer tokens.