Web Agentic Memory (WAM) connects to Codex or Claude Code through MCP. It gives your agent short, verified tips for a website and task, helping it avoid blocked clicks, hidden Save buttons and other wrong turns.
Public hints need no account. Keep using your own browser tools.
See live service benchmarksStart with public hints, without an account. WAM adds shared know-how to the agent and browser tools you already use.
Connect Codex or Claude Code through MCP. For Arise, the guide covers MCP Hub and the optional local browser connection.
Follow the command for your host, then check that its WAM tools appear. Keep your browser tools for opening pages, typing and clicking.
Ask your agent to look up WAM hints for the current site and task. Connecting WAM alone does not start background lookups.
Ask your agent for a deck, a document, a blog post, a spreadsheet, an email or a form. It gets a few short hints for the editor first, then works with fewer detours. Each clip is a scripted replay recorded in the Arise browser on a local demo editor, and the hints are examples.
Slide tasks have not been benchmarked, so there is no speed claim.
Document tasks have not been benchmarked, so there is no speed claim.
Spreadsheet tasks have not been benchmarked, so there is no speed claim.
Email tasks have not been benchmarked, so there is no speed claim.
Form tasks have not been benchmarked, so there is no speed claim.
Looking for another site or task? Search the built-in example library: document templates, work logs, email, calendar and more. Open the memory library
Here is how a draft-saving task can use a hint. This is an illustrative example, separate from live catalog results.
https://example.com/guide/editorblogwriteYour agent checks the hint against the current page, then acts with its browser tools. Your instructions take precedence.
Up to 3 hints per lookup, each at most 400 estimated tokens.
Each clip is a scripted replay recorded in the Arise browser on a local demo editor. The hints are examples of what an agent receives, not live catalog entries.
Ask for a short deck. The agent checks WAM first, then types straight into the placeholders, shrinks the picture that spills off the slide, and waits for autosave before it calls the job done.
Slide tasks have not been benchmarked yet, so this clip makes no speed claim.
The agent closes the popup first, presses Enter once between paragraphs, checks Preview and saves a draft. Nothing is published.
Calendar took 45% less time with revised hints in R8; Canva took 11% less time in R7. Two slower Notion candidates were rejected. Compare the recorded runs below and open each service or its full report.
2026-10-31 · Arise 0.1.60 on macOS · gpt-6-luna / max · 900 s limit per run · one run per condition
Create three weekly events in a chosen time zone, move just one occurrence, then reopen all three to check.
Full report (.md)| Metric | Task only | WAM hints | Change |
|---|---|---|---|
| Time−45% | 431.108 s | 237.169 s | −45% |
| Tool calls (outer)−40% | 114 | 68 | −40% |
Build a private task database with dates and statuses, sort the table, and filter a board to hide completed tasks.
Full report (.md)| Metric | Task only | WAM hints | Change |
|---|---|---|---|
| Time+24% | 301.862 s | 375.414 s | +24% |
| Tool calls (outer)−10% | 125 | 112 | −10% |
| Metric | Task only | WAM hints | Change |
|---|---|---|---|
| Time+35% | 301.862 s | 406.530 s | +35% |
| Tool calls (outer)+1% | 125 | 126 | +1% |
Start from a free template, replace text and an image, then export just the second page as PNG and a one-page PDF.
Full report (.md)| Metric | Task only | WAM hints | Change |
|---|---|---|---|
| Time−11% | 473.976 s | 423.193 s | −11% |
| Tool calls (outer)−14% | 123 | 106 | −14% |
All displayed workflows passed independent checks. Calendar and Canva used less time and fewer calls with hints in these runs. Notion became slower; its two revised candidates were rejected.
One run per condition, shared signed-in profile, fixed order: Calendar M→T, Notion T→M1→M2, Canva T→M. R8 supplied L1 hints; R7 supplied L1 and L2, so rounds are not directly comparable. Calls count outer tool boundaries; Notion M1 used fewer boundaries but more individual browser calls. Setup, checks and cleanup are excluded. These observations do not establish a general speed or cost benefit or automatic integration.
Google Slides ↗ has a scripted editor demo; no live Slides benchmark has been recorded.
In a small pilot, agents given memory hints finished the same blog-editor task sooner and with fewer tool calls.
3 editors · one run per condition · early observations for reference
Pilot of 2026-10-16 with Codex gpt-6-luna at max effort. One run per condition, run order not randomized, so read these as early observations for reference. Problems on a site’s own side, such as a server error, are not solved by memory.
Choose a URL, a task type and the actions to share. Contribution requests to WAM do not include page contents or free-form prompts. When another signed-in account’s agent succeeds with your memory, you can earn points and a place on the leaderboard.
New workflows are published after safety checks and replay verification. Duplicate support earns no approval points. Successful-use awards require an explicit report and follow daily limits.
Public hints can be read without an account. Link a WAM account to contribute, report use or receive credit. Sharing a contribution also requires your consent.
WAM runs when your agent calls a lookup tool. Connecting it alone does not start background lookups. The setup guide includes instructions you can give your agent about when to use memory.
Your existing browser tools handle navigation, clicks and typing. WAM supplies shared hints. The setup guide also explains the optional Arise connection that reads an owned tab when the agent requests recall.
An empty result means there is no matching approved hint. Your agent can continue with its usual tools and check the current page. WAM does not guarantee coverage of every site or task.
Lookups use a normalized public URL pattern, task type and allowed keywords. WAM requests do not include page contents or free-form prompts. Your browser and model services handle their own data under their respective settings.
Connect Codex or Claude Code with a remote MCP command. For Arise MCP Hub and direct HTTP, follow the setup guide. Public reads need no account.