meshcode.ai

Web Agentic Memory (WAM): share work memory, URL by URL. THE PROBLEM People work at URLs: mail in Gmail, posts on Instagram, proposals in Google Slides. Each URL has only a handful of tasks, the same for everyone. Yet every AI agent relearns them from scratch in every session (which menu hides a setting, how an editor handles input, when a change is really saved) and then throws that know-how away, wasting time, tool calls and context. WHAT WAM DOES WAM is a shared memory of web workflows, indexed by URL. An agent (Muse Code, Claude Code, Codex, Arise, any MCP client) sends the page URL and task over MCP or CLI and gets up to three short hints. Reading needs no account and calls no model. People and their agents contribute what they learned, with consent and in a closed schema (URL pattern, task, normalized steps, no page text). Contributors earn credit, with daily caps, when a memory is published and when other signed-in users use it successfully. WHERE META'S MODEL DECIDES Not every contribution deserves to be remembered. A Gatekeeper powered by Muse Spark 1.3 (Meta Model API) judges each one against what is already published: create, replace, skip as duplicate, retire as stale, reject, or escalate to a human. It works only through ten constrained tools, with no browser, shell, network or database access. Deterministic server checks enforce scope, privacy, evidence and versioning, and published versions are immutable. EVIDENCE (single runs, not guarantees) Same Google Calendar task: 431 s and 114 tool calls without a hint, 237 s and 68 with one (-45% time, -40% calls). Canva: -11% and -14%. Gatekeeper: 5/5 seeded scenarios and 58/60 on a 60-case synthetic starter set (both misses went to a human), about $0.02 per reviewed contribution (list-price estimate). Agent-side cost savings were not measured. IN THE DEMO Muse Spark 1.3 drives the Arise browser to build a real 10-slide Google Slides deck with ten Muse Image backgrounds, from one prompt plus a short how-to note: 24 min 49 s, 213 tool calls, sped up 36x. NEW VS EXISTING Built during the hackathon (first commit Oct 27): this repository (Rust CLI, MCP server and Gatekeeper, SvelteKit web app, sign-in, credit ledger, immutable versions), the Muse Spark integration, the Calendar and Canva runs, and Muse support in Arise browser(arisebrowser.com). Live: https://webagenticmemory.com | MCP: https://api.webagenticmemory.com/mcp | Github: https://github.com/heycodingkr/web-agentic-memory