47 Agents All Need Desktop Permissions; OpenClaw Puts the Memory Problem on the Table First

OpenClaw posted a very short main tweet: the number of agents keeps growing, but there is no need to keep everything in memory—we have a way. The attached image is McBain’s weekly roundup of Personal AI Agents.

The tweet itself doesn’t contain much information; the real information is in the quote thread. McBain listed the week’s updates from products including OpenAI, SpaceXSI, Meta, Anthropic, OpenClaw, Nous Research, Perplexity, Instinct, Pally, Wabi, and Strawberry. The comments then laid out another side of things.

Original link: https://x.com/openclaw/status/2115905733351120900

This Week, the Big Players Are All Fighting for the Entry Point

OpenAI’s dots connected to iCloud Mail and Calendar plugins on October 27, covering all plans across web, desktop, and iOS. On October 28, dot can send and receive text, photos, and voice via iMessage or RCS, and will also proactively push updates. US Pro users get the preview first.

Grok Bot comes from @SpaceXSI. On October 22, it can use phone apps such as Calendar, Reminders, and Contacts, and can also log in with 1Password. On October 23, it saves login information, cards, and contacts. On October 26, it can import memory from another AI assistant. On October 27, it added shopping cards, order tracking, and Shop Pay.

Meta’s Muse launched in Japan on October 22; on October 23, v13.0 added iPhone Duo support, a Feed home-screen widget, and in-chat shopping; on October 27, it launched in Singapore. Version information comes from whatsnew.fyi’s tracking.

Anthropic’s Claude Duet published a more detailed safety, privacy, and safeguards note on October 23, without releasing new features.

These updates share a common direction: bringing agents closer to your entry points. SMS, iMessage, calendars, contacts, logins, shopping, memory import. They no longer just wait in a chat box for your questions; they’ve gained a place in the phone’s operating system.

Open Source and Self-Hosted Haven’t Stopped

OpenClaw shipped three releases this week. On October 24, v2026.10.7 fixed update checks, Claude CLI, and Bun. On October 27, v2026.10.8 fixed installation, updates, and stuck chats. On October 29, v2026.10.9 delivered security fixes. On October 26, it integrated GPT-6.1 Terra, with login via ChatGPT or an API key. On October 28, the web UI gained offline support, per @steipete.

Nous Research’s Hermes Agent added the Advanced Delegation plugin on October 26, allowing a separate model to be chosen for each subagent, while also integrating GPT-6.1 Terra via Nous Portal and OpenRouter. On October 27, new plans all include a cloud agent with daily and weekly limits. On the same day, the desktop app added ElevenLabs sound effects.

ZeroClaw released v0.8.7 on October 27, fixing Wasmtime 48.0.7 and TLS verification, handling Telegram rate limits, adding ZeroCode cost warnings, and allowing chats to reconnect after refresh. This release had 81 commits from 19 contributors.

nanobot released 0.4.0 on October 28, with more reliable chat history, the ability to connect to remote instances, Mermaid diagrams rendered in the terminal, and a smoother mobile experience.

McBain also left a section for independent products: Wabi became a unicorn, Instinct got its own spending allowance, Pally ships almost every day, and Strawberry reached 1.0.

The Comments Are Where the Real Action Is

One reply said they were relieved—there are already too many agents to keep track of. Someone asked what these things are actually useful for to business owners. Another asked how long it will be until there’s a fully open-source JARVIS that appears on screen as an avatar.

A more specific one came from ShadowAguy: 47 personal agents, each wanting full desktop permissions.

Debenhams’s long reply was even more blunt. He said managing the new secrets only in the Control UI is agonizing, memory has also had problems, the agent doesn’t remember tokens, and a few hours later it goes around things the old way again. Workflows that used to be simple now take days to troubleshoot. He said his confidence in OpenClaw is dropping day by day.

Shenpiao Chuanzai asked about the technical underpinnings: after running a few more agents, memory overhead quickly spirals out of control—how exactly is state offloading handled? draslan, meanwhile, wants to use only OpenClaw to run a software company.

Put these two kinds of replies together, and the problem is clear. Everyone wants JARVIS, but first they have to get past 47 agents’ desktop permission pop-ups. Every agent wants permissions, wants to remember state, and wants to get keys. Once the number grows, memory and secrets turn from backend details into everyday failures.

OpenClaw’s main tweet says you don’t need to keep everything in memory. To developers, that sentence is about state management; to users, it’s about why it forgot the token again. What’s missing in between is productization. What’s really worth watching this week isn’t who integrated another entry point, but that as agents move from the chat box toward system permissions, memory and key management are starting to become part of the product experience. Whoever can hide this dirty work well is the one who deserves to talk about personal agents.