AI Prompt
BlogAbout
← Back

Anthropic on Agent Memory, OpenAI's BioBasicsBench, and More

4 min read

Anthropic published a useful piece on how agents should remember things between sessions. It splits memory into four layers: transcripts (a searchable log of what happened), notes the agent writes as it works, structured representations like wikis, and context that gets injected into every session. On top of that sits "dreaming," an offline pass where Claude reviews transcripts and existing memory, fixes mistakes, and reorganizes what it knows before the next session.

The numbers are what make it worth reading. Anthropic built a wiki from about 350,000 files at Stainless, roughly 420M tokens. Using only the wiki, the agent hit 13% of graded criteria after 2 dreaming sessions and 48% after 128. Wiki plus search reached 61%, compared with 53% for search alone, and it saved about 20,000 tokens per question.

The part I keep thinking about is the failure mode. When several agents wrote to one shared file at the same time, they overwrote each other and facts got lost. Giving each agent its own page fixed it. If you're building multi-agent systems with shared memory, that's the lesson to take away. Anthropic.

What Else Happened

OpenAI released BioBasicsBench, a benchmark for everyday wet-lab analysis. It was built with Ligo, an enzyme-design company, and has 100 tasks taken from real experiments: gels, colony plates, microscopy, assays, and cloning. GPT-6 Astra scored about 66.7% mean credit at its highest reasoning setting, compared with about 43.8% for GPT-5.6 Sol. It's still weak on controls, imaging, and reading context. I like that it tests whether models can read messy lab data, not just answer textbook questions. (OpenAI's post, X story)

Google DeepMind says human-guided AI agents found two weaknesses in legacy 1024-bit RSA. One is about forging signatures and the other is about factoring 1024-bit keys in bulk more cheaply. DeepMind says 2048-bit and longer keys aren't affected. If anything still relies on 1024-bit keys, like old internal CAs or DKIM records, now is a good time to replace them. (Google DeepMind, Pushmeet Kohli's thread)

Google found that stale TODO comments send coding agents on detours. Across its internal agent sessions, 13% included inline TODO(bug/...) or FIXME(bug/...) references. Agents ignored most of them, but sessions where they went chasing a bug used about 77% more tokens and 105% more wait time. Many of those bugs were already closed. Google then had agents clean up thousands of obsolete references, and human reviewers accepted over 95% of the changes. Comments are now part of the prompt, so they need to stay current. (Google Developers Blog)

GitHub detailed a new Git storage architecture built for agents. It separates storage from serving. In internal tests, GitHub says it handled 5,000 concurrent agents sending 7,200 pushes per second to one repository. Those are GitHub's own test numbers, but it's clear where this is going: many agents working on the same repo at the same time. (GitHub Blog)

Apiiro found secrets in 1,122 public copies of Claude Code's .claude/settings.local.json. When you pick "Yes, and don't ask again," Claude Code can save the exact command as an allow rule, including any token or password in it. Counting git history, shared settings, and transcripts, Apiiro found 2,539 secrets across 1,994 repositories. If this applies to you, rotate the keys first, then stop tracking the file and clean your history. Deleting the file alone doesn't help, because the old versions stay in history and forks. (The Hacker News)

Claude Code cloud sessions now set up their own environments. On Pro, Max, and Ultra, Claude figures out what dependencies, network access, and API keys a repo needs, then saves the environment for next time. Keys are stored as network secrets, so Claude never sees them. That's the right design, especially given the story above. (Noah Zweben's post, cloud environments docs)

Launches and Demos

  • OpenCode Browse – an MCP server plus a Chromium extension that lets agents use your real local browser, with your tabs and logins. When a login, 2FA prompt, or payment comes up, it hands the tab back to you. It works with OpenCode, Claude Code, Cursor, Codex, and Zed. (X story)

What I'm Reading

That's it for today. 👋