OpenClaw Releases Its Largest Update Ever: 2.0 Rebuilds Memory, Session, and Permission Systems

After nearly seven weeks of silence, OpenClaw—the open-source AI agent project that once took the world by storm—officially released version v2026.8.1 on August 31, which the team is calling OpenClaw 2.0. This is the largest update since the project's inception, with 933 contributors participating, 569 of whom submitted code for the first time, and over 16,000 pull requests merged—nearly half of the project's total historical merge count.

The release blog post, written by team member Hannes Rudolph, carries a telling title: "OpenClaw 2.0, Accidentally." According to Rudolph, the team initially set out to do just two things: simplify the installation process and make the browser experience a "first-class citizen." But as they pushed forward, they discovered that merely swapping out the interface layer was far from sufficient. Compatibility issues had accumulated across modules like Session, Memory, Plugin, Automation, and Gateway, and existing users could still encounter problems when migrating data. Ultimately, the team rebuilt nearly everything—installation, messaging, memory, Skills, model management, Automation, browser, native client, Plugin, security, and Session systems—even overhauling the release process itself.

Memory System: From Session-Level to Long-Term Retention

In the 1.0 era, OpenClaw's context would easily collapse once it grew too large, forcing users to re-explain their background every few days. 2.0 introduces an Active Memory mechanism that allows the agent to pull relevant context from past conversations when conditions are met in a given session.

The newly added Background Memory Consolidation feature lets the model automatically organize long-term memory in the background, promoting genuinely valuable information to long-term storage while preserving source attribution and a "Dream Diary." The Dream Diary refers to the agent's off-hours process of organizing memories accumulated during work.

Going a step further is the Automatic Self-learning mechanism. After completing a task, the agent can summarize reusable experience, pass it through safety checks, and either create a new Skill or improve an existing one—achieving a degree of self-evolution. Project founder Peter Steinberger revealed that the new version deleted a redundant Memory system, cutting roughly 50,000 lines of code and keeping only the unified architecture. He also admitted candidly: "Nobody has really solved the Memory problem yet. What OpenClaw does now is separate 'what happened in the past' from 'what's worth remembering long-term.'"

Session System: From Single-Machine Limitations to Cloud Collaboration

The most striking product change in 2.0 is Shared Cloud Sessions. Multiple people can join the same task, handing off work with full context intact, with permission tiers divided into four levels: read, suggest, draft, and direct participation. Rudolph said the feature emerged from the team's own experience: as they increasingly delegated work to their own Claw during 2.0 development, they realized there was no way to bring another person into a task without losing context.

According to the team, they now use this very feature to develop OpenClaw itself. Steinberger described it this way: "Multiple people coding together, combined with the near-unlimited compute from nodes and cloud Sessions, has completely changed how we do development." He even suggested that traditional local coding harnesses now feel "like relics of the past."

In the new version, Sessions can be moved from the Gateway to a paired computer or a Cloud Worker, with workspace and context migrating together. Cloud machines shut down when idle and are restored when needed. Steinberger's vision: the Gateway serves as an always-online brain, while computers, phones, and temporary cloud servers are merely Nodes that can be dynamically invoked.

Permission System: From Loose Authorization to Granular Control

Security changes are equally sweeping. 2.0 adds a Private Credential Request mechanism, where passwords and API keys are entered through a masked window that never enters the chat history or the model's context window. Automation now features more granular approval: fixed tasks can receive persistent authorization, but if the task or operation changes, re-confirmation is required. Plugin installation displays source, version, and specific capabilities, with plugins from unknown sources that can execute code requiring additional approval.

The team also published crowdsourced red-team data: across 272,000 attacks in 41 agent scenarios, using "executed a harmful action while deceiving the user" as the success criterion, the breach rate was only 0.5% with Claude Opus 4.5, 1.0% with Sonnet 4.5, 1.3% with Haiku 4.5, and 8.5% with the worst performer, Gemini 2.5 Pro. However, the documentation explicitly notes that against adaptive human attackers, even the best existing defenses are breached more than 80% of the time—tool policies, execution approvals, and sandboxes are the real enforcement layer.

Installation and Performance: From Tinkerer's Toy to Out-of-the-Box

In the 1.0 era, the barrier to raising a "lobster" was absurdly high—API keys plus a pile of command-line configuration meant anyone without sufficient computer knowledge was essentially locked out. 2.0 simplifies first-time installation to an out-of-the-box experience. The new guided installer first scans the machine for existing AI access permissions, can reuse already-verified Codex, ChatGPT, or Claude CLI logins, accepts API keys, or discovers locally installed Ollama and LM Studio models. The browser client has been completely rewritten, with the old Overview page removed—startup to conversation now takes just 575 milliseconds. The underlying storage has migrated from the filesystem to SQLite, at the cost of no longer offering free downgrades—the team recommends performing a verified backup before upgrading.

A Difficult Pivot After the Hype Fades

The timing of 2.0's release coincides with a dramatic decline in OpenClaw's popularity. According to data from Fanfan Research, OpenClaw's GitHub traffic fell from a peak of approximately 28.8 million visits in April to 14.2 million, and had dropped below 10 million before this update. Weekly new GitHub stars fell from an average of 5,900 in February and March to roughly 800 in the most recent two weeks. By comparison, rival Hermes Agent is adding nearly 9,000 stars per week—a gap of more than 10 times. NPM weekly downloads fell from a mid-March peak of over 1.5 million back to pre-viral levels.

Reflecting on the experience on YC's startup podcast, Steinberg admitted: "Be careful what you wish for. I wasn't prepared for all this attention—it nearly crushed me." He revealed that his private phone number was leaked, and journalists would call late at night using emergency breakthrough mode. After joining OpenAI, he found himself waking up with two jobs: "one is a job, and one is a mission." The worst symptom: he stopped using OpenClaw himself.

Upstream supply disruption also contributed to OpenClaw's decline. In February, Anthropic updated its terms to prohibit connecting Claude subscription OAuth credits to third-party tools like OpenClaw. On April 4, Anthropic formally emailed users that Claude subscriptions would no longer cover usage through third-party tools like OpenClaw. Previously, the vast majority of users were using Claude to power their "lobsters"—this move doubled their costs.

On the podcast, Steinberg admitted to a strategic error: "To make OpenClaw perform at its best, I deeply optimized the system for Anthropic's models—and that company ended up becoming my competitor." He also delivered a line that almost amounts to a self-deconstruction of the "lobster myth": "Eight months isn't long in human time, but in AI time, it's equivalent to four years."

Competitive Landscape: Surrounded by Rivals

The features OpenClaw 2.0 is shipping have long existed in other agent products on the market. Nous Research's Hermes Agent went open-source in late February, touting self-evolution, four-layer memory, built-in sandboxing, and dangerous command approval—it racked up tens of thousands of stars in two months, with the community rallying around the slogan "from raising lobsters to raising horses." Moonshot AI released KimiClaw, fully cloud-hosted with zero-barrier registration; Alibaba Cloud's Tongyi launched CoPaw, deeply integrated with DingTalk; Tencent has QClaw, and ByteDance has Moltbook.

The Crabbox project Steinberg developed after joining OpenAI has only about 1,300 stars on GitHub—less than OpenClaw's daily growth even during its decline. However, his methodology lives on in Codex. Codex introduced multi-agent parallelism with long task loops, a mechanism Steinberg originally proposed in OpenClaw. In the first half of 2026, Codex's active users grew more than 5x, reaching 20 million by August.

Community reaction to 2.0 has been noticeably warmer than recent updates, with Multiplayer, Self-learning, and Cloud Session drawing the most attention. Core contributor Vincent Koc called it "the biggest lobster ever." But familiar problems haven't fully disappeared—within hours of release, Reddit users were already reporting "I let it upgrade itself, and then it died," while others said OpenClaw entered a boot loop on M4 Pro after updating, pegging CPU at 100%. The team has acknowledged discovering migration issues and is preparing further patches.

Steinberger is transforming OpenClaw from a "chatbot that can control your computer" into a persistent computing layer: the Gateway stays online continuously, retains Memory, dynamically orchestrates computer and cloud resources, and spawns or invokes different Agents as needed. The fundamental unit of the future Agent may no longer be a prompt or a chat session, but a persistent software entity with memory, permissions, tools, and compute resources. Whether the "lobster" can regain its edge will ultimately be decided by the most practical test of all.