← Blog
Models

Your agent doesn't forget last month. Old conversations get filed, not deleted.

An AI model can only hold so much in its head at once. Most tools handle that by dropping the oldest messages, which is how an agent that knew your process in March forgets it in April.

Bench2 min read

The short version

  1. AI models can only hold so much at once. Most tools drop the oldest messages when they fill up.
  2. On a Bench harness, old messages are saved and summarized, never deleted.
  3. The agent can search back and open the original text.
  4. The software is lossless-claw by Martian-Engineering. We run it. We credit them.

The problem

Every AI model has a limit on how much it can look at in one go, like a desk that only fits so many pages. When the desk fills up, most agent software sweeps the oldest pages into the trash. The agent keeps working, but it no longer remembers the thing you told it three weeks ago about how your office handles a denied claim. It feels like training a new hire every Monday.

What we run instead

On a Bench harness, nothing gets swept into the trash. Every message is saved to a database on your Mac. When the desk fills, older pages get summarized, and the summary points back to the original. The agent works from the summaries and can open the original whenever it needs the detail. So it can still find what you said in March, word for word.

Who built it

We didn't invent this. The piece of software that does it is called lossless-claw, built by a team called Martian-Engineering from a published research paper. We run our own version of it on every harness, and we say so, because it's their work and it's good.

What it means for you

The rules you teach your agent stay taught. That's most of the difference between an agent and a chat window you have to re-explain yourself to.

Questions

Do old messages get deleted?

No. They're saved in a database on your Mac. Summaries point back to them, and the agent can open the original.

Did Bench build this?

No. Martian-Engineering built lossless-claw from a published paper. We run our own version and credit them.

Why does this matter?

An agent that forgets what you taught it last month is a chatbot again. This is what keeps it an agent.

Keep going

Every Bench word above is on Glossary / FAQs, in plain English.

New posts

We'll email you when a new post lands.

Email signup opens soon. Until then, new posts show up right here.

More from the blog

Blog

Back to all posts

How Bench works. Notes on the models, the software, and the jobs we run.

Next step

See Bench on your pipeline.

You approve everything. Your data is yours — even if you leave. $249/mo, unlimited users.