Home / Vault & Second Brain / Anthropic Named What My Agent Already Does

Anthropic Named What My Agent Already Does

Abstract second brain dreaming its memories into files

Anthropic Named What My Agent Already Does

This morning I read that Anthropic is turning memory into the next primitive for their agents. They gave it a name I couldn’t stop thinking about: dreaming. Not the romantic kind — the practical kind. At night, an agent takes yesterday’s logs and quietly turns them into new memory and new skills. I read that and had to smile, because I have been running exactly that for weeks.

Every night, a cron job in my setup reads through what happened during the day — the runs, the errors, the little corrections — and consolidates them into something durable. It is my own Mnemosyne, a vector memory, plus a layer called Hindsight that does retain, recall, and reflect. The loop is the same one Anthropic describes: experience → remember → connect → retrieve → act → update. I just never had a clean word for the consolidation step. Now I do. It’s dreaming.


Memory as a filesystem, not a database

The second idea in their announcement is the one that genuinely matched a decision I made on my own. Anthropic presents memory to their model as a filesystem — a hierarchy of files the agent itself manages with bash and grep. No hidden API, no magic. Just files the agent can read, edit, and own. That is, almost word for word, how I keep my own second brain: everything in the vault as plain Markdown files, versioned with git, with the provenance visible in every line. It turns out I didn’t invent something eccentric — I drifted toward the same shore as the frontier.

But here is where the announcement stung a little. Anthropic demands three properties for a memory layer, and I only have half of one. They want permission scopes (read-only knowledge versus read-write working memory), optimistic concurrency (a content hash so you know whether another agent overwrote your file), and a standalone portable API so you are never locked in. My setup has the portable part — it is just files and git, nothing proprietary. The other two, I do not have. If two of my agents touch the same note at the same time, I would not notice until the history is gone.


The real lesson is about verification

What struck me most, though, is a smaller detail buried in all of this. My system has a habit of declaring things “done” when they only look done — a monitor that says a backup is valid while the file never landed, an uptime number that says 100% while the newest real work is months old. Anthropic’s answer is a content hash you check before you overwrite anything. My answer so far has been a verification gate after the fact. Their way is better: the check lives inside the write, not after it.

So the question I am sitting with today is not “should I adopt Anthropic’s stack?” — I already have my own. The question is: which of those three properties am I missing most, and is it worth building the check into the write instead of bolting it on after? That feels like the difference between a system that remembers, and a system that can prove it remembers.

Do you treat your notes as a database or as files you own? I’d genuinely like to know — the comments are open.

Wat vond je van dit bericht?

Leave a Reply

Your email address will not be published. Required fields are marked *