Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

After an Anthropic ban, Codex moved 43 Claude Code sessions in 20 minutes

A developer was banned by Anthropic after six months of daily work with Claude Code, but his working context survived: the agent's rules and memory lived outside the account. Codex, a competing agent, imported his 43 latest sessions, skills and hooks in 20 minutes. By evening, Claude on a personal server already knew the project's 19 rules. The practical lesson: keep your agent's rules in files and in git, not in the settings of a single service.

AI-processed from Habr AI; edited by Hamidun News
After an Anthropic ban, Codex moved 43 Claude Code sessions in 20 minutes
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1claude.ai chats vanished along with the account, while Claude Code sessions stayed on disk
Key metric #2Claude Code silently dropped 16 of 35 memory tools with complex schemas
Key metric #3The "no comments in tasks" rule failed because no general rule had been recorded

Imagine learning on a Monday morning that your AI provider has blocked your account, and that six months of agent setup seem to have vanished with it: who gets assigned merge requests, which server is where, which rules you spent months explaining. That is how the morning of November 2 began for the author of this story, a member of the Lamantin AI lab. Anthropic banned him even though he worked in Claude Code every day: code, tracker tasks, document analysis, writing.

But within 20 minutes of starting the migration, Codex (OpenAI's competing agent) had moved his 43 latest Claude Code sessions along with skills and hooks, and by evening he was back at work without re-explaining anything. Below, we look at why it went this way and what to set up in advance so a ban doesn't become a catastrophe.

Where an agent's knowledge actually lives

By the time of the ban, the author had accumulated more than two hundred Claude Code sessions, plus his own skills, hooks and rules. He split what the agent "knows" into four sources, and each fares differently when an account is blocked.

The first source is conversation history. Chats on claude.ai are tied to the account and disappear with it. Claude Code sessions work differently: they are ordinary files in the ~/.claude/projects folder on your computer, and the ban did not touch them.

The second is project instructions: CLAUDE.md, AGENTS.md and similar files. They live in repositories, so nothing happens to them. Each tool does have its own format and location, though.

The third is skills and hooks. These are also files, but configured for a specific tool.

The fourth is memory: rules, decisions, facts about servers and people, open tasks. The author keeps it in kaeru, a free memory tool for agents built by his lab. It is a database on his computer that agents connect to over MCP, a common protocol for agent tools. Claude Code, Codex, Cursor and others work with it natively. It doesn't matter which agent wrote a rule: any connected agent can read it.

How the migration went

At 9:26, the author asked Codex to pull in his sessions from Claude. Codex has a built-in import from Claude Code. It transferred the 43 latest conversations as regular chats, along with skills, hooks and the connection to kaeru. Codex put the remaining sessions into an archive and created a separate skill for searching them. The whole thing took about 20 minutes.

It was not entirely smooth: the import rewrote the paths to skills incorrectly and pulled in Claude's internal system instructions, and the context-compaction hooks had to be adapted to Codex's format. Still, at 9:51 the author opened a new chat and told it to continue the "vacuum cleaner" session, where he had been working through work tasks. Codex found it, pulled the project's rules and state from kaeru, checked the tracker and proposed the next task.

That evening the author decided to bring Claude back, not on his own computer but on a personal server under a new subscription. The reason is pragmatic: the "disguise" could suddenly stop working, so it is safer to put the agent on a fixed IP address. Codex set up the server.

It brought up a tunnel to the computer holding the memory, moved the skills over and assembled a CLAUDE.md from the same rules. At 22:16 the author asked the new agent whether it had access to memory.

It replied that at startup it had loaded 19 rules, one open task, one entry awaiting review and another 316 entries that can be pulled up on request. Some of the rules had been written by the old Claude in October, and some by Codex the same day.

The next morning, the agent said that before starting work it would first assign the task to the author, set the status to "in progress", reread the card and only then go into the code. Codex had written that rule the day before. The author adds an honest caveat: the rule reached the agent by two paths, from kaeru and from CLAUDE.md.

Where the migration stumbled

Problem one: on the first evening, the agent on the server was missing 16 of kaeru's 35 tools, including search, reading and writing notes. The agent decided they had been disabled in the settings, but enabling them did not help. Codex found the cause: these tools have complex parameter schemas, and Claude Code silently dropped them. Codex wrote an intermediate layer that simplifies the schemas, and after reconnecting everything worked.

Problem two: memory remembers only what was written down. In October the author had asked the agent more than once not to write comments in tracker tasks until he gave permission, but this ended up in notes on individual tasks, with no general rule recorded. On November 3, the first thing the new agent did was leave a comment in a task.

Problem three: a stale entry leads you astray. Codex answered from memory that Antigravity had not worked on the server the day before because of the account's region. The author doubted this, sent a screenshot of an email from Google, and the entry was updated. The memory moved over with its errors intact.

Problem four: the server depends on the computer where the memory lives. Turn the computer off, and the agent is left without it.

What to do in practice

The principle works without kaeru: keep what the agent knows about you outside the account and outside any single tool. Store working rules in files that you hold and keep in git, not only in chat settings. If an agent saves conversations locally, as Claude Code does, back up that folder. Turn a repeated request into a rule right away, not on the eighth time: a note saying "no comments on this task" will not help a new agent.

For those who want to try kaeru: install the daemon, connect it to each agent (for Claude Code, with the command claude mcp add --transport http kaeru http://127.0.0.1:9876/mcp) and add a hook that sends the agent to memory at the start of a session.

The author's bottom line: he regrets the subscription money, but the switch turned out to be simple. The main asset when working with agents is the accumulated rules and context, and you should manage them yourself.

Story

Key Questions Answered

Where an agent's knowledge actually lives?

By the time of the ban, the author had accumulated more than two hundred Claude Code sessions, plus his own skills, hooks and rules. He split what the agent "knows" into four sources, and each fares differently when an account is blocked. The first source is conversation history. Chats on claude.ai are tied to the account and disappear with it. Claude Code sessions work differently: they are ordinary files in the ~/.claude/projects folder on your computer, and the ban did not touch them.

How the migration went?

At 9:26, the author asked Codex to pull in his sessions from Claude. Codex has a built-in import from Claude Code. It transferred the 43 latest conversations as regular chats, along with skills, hooks and the connection to kaeru. Codex put the remaining sessions into an archive and created a separate skill for searching them. The whole thing took about 20 minutes.

Where the migration stumbled?

Problem one: on the first evening, the agent on the server was missing 16 of kaeru's 35 tools, including search, reading and writing notes. The agent decided they had been disabled in the settings, but enabling them did not help. Codex found the cause: these tools have complex parameter schemas, and Claude Code silently dropped them. Codex wrote an intermediate layer that simplifies the schemas, and after reconnecting everything worked. Problem two: memory remembers only what was written down.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…