AI Agent Notes - (3)

Recently, I have been thinking that an agent is not only the LLM.

The model is like the brain core, and the settings shape its behavior, but a large part of the real working brain is in memory and skill. If we want to really deliver an agent from one place to another, those parts should be transplanted too.

That is why I like this small tool:

https://github.com/hyharry/agent-brain-transplant

A simple picture is:

1
2
3
4
5
6
7
8
9
10
11
12

[LLM+settings]
      |
      v
[memory+skills]
      |
      v
 [real agent]
      |
 transplant
      |
      v
 [new place]

It points in the right direction. If an agent has built useful habits, workflows, and stored context, moving only the model is not enough.

What still feels missing is benchmarking. It is not hard to say an agent has skills and memory, but it is more important to test whether those skills are actually used and whether the right memories are really retrieved. If someone wants to make full use of an agent, useful benchmarks should be part of the process.

Another missing piece is imagination and creation. Humans do not only retrieve memory. We also connect distant ideas, or sometimes create something that seems to come from nowhere. Part of that may be approximated by finding hidden connections inside the agent brain through memory and skill retrieval. But imagination still feels different.

For that, I think OpenClaw creator Peter Steinberg’s idea of daydreaming is a good starting point.

Maybe the next step for agents is not only better memory, better skill, or better benchmarks, but also better ways to wander before returning with a new idea.

AI, OpenClaw, agent, memory — Jul 15, 2026