AI agent memory: what it is and how Loomfield remembers how you work

What is AI agent memory?

AI agent memory is the information an assistant carries from one step or session into the next. Three kinds are commonly distinguished. Short-term memory is the current conversation, which the model can see while you work. Long-term memory is stored information, such as facts and preferences, that is brought back in later sessions. Procedural memory is knowledge of how to do something: which sources to open, which checks to run, how the result should be shaped.

Most memory features focus on the first two. For recurring work, the third matters most, because it is the part you otherwise rebuild by hand every time.

Why does AI forget how you work?

A new chat starts with an empty context. Unless something deliberately carries information forward, you find the files again, explain the business again, and rebuild the checks that made the last answer trustworthy.

The answer itself may be saved in the old conversation, but the method that produced it is scattered across prompts, corrections, and decisions you made along the way. That is why a good result in one chat does not make a related job in another chat any easier.

What does Loomfield remember?

Loomfield is designed to keep how a job was done, not only what was said. That includes the sources that mattered, the checks that were run, the choices you made, the corrections you gave, and the format you settled on. These stay connected to the finished result.

The aim is to reuse the method without forcing every new job into the shape of the last one. A monthly report can teach a later variance question which sources matter, while a different job still gets its own judgment. You approve what carries forward.

Status: Loomfield is in private beta. Learning from finished work and proposing reuse is the product direction and is in development. See how Loomfield works for the full picture.

Where does Loomfield keep its memory?

In the Weave, your private knowledge base. The Weave lives in your home folder on your own computer. What you and your sessions create is saved there.

Your computer acts as the hub, so projects, work history, methods, and AI connections stay in one place instead of being spread across separate chats and services. When Claude or Codex works on a task, the content it is given goes to that provider under your own plan, the same as it would with their own tools. See the glossary for the Weave, hub, and project terms.

How does remembered work get reused?

Loomfield describes one learning loop in four steps. First, do real work with AI helping alongside you. Second, keep what made it good: sources, checks, tools, choices, corrections, and formatting. Third, connect related work, so a new report can inherit a proven analysis method. Fourth, automate only what becomes stable.

Automation is the last step and stays your decision. Work can stop at any level, from AI that helps when you ask, to prompts when work needs attention, to prepared drafts that wait for your approval, to routines that run within boundaries you approved.

How is this different from chat memory?

Chat memory typically stores things you told the assistant. Loomfield aims at how your work gets done and keeps it under your control on your machine, beneath the AI tool you choose. The comparison is covered in detail in persistent memory for Claude and Codex.

Common questions

Does Loomfield work with Claude and Codex?

Yes. Loomfield uses your own login with a Claude Pro or Max subscription, or a ChatGPT plan with Codex. It never asks for an API key.

Is Loomfield available now?

It is in private beta. You can request access from the homepage, and the getting-started guide walks through setup for Mac and Windows.

Do I need to build a workflow first?

No. You begin with a real job. The point is that doing real work builds the system, so there is no workflow builder to learn first.