A field guide to agent memory

A small thought.
A much bigger
possibility.

Your mind can hold a few things at once. Your notebook can hold a lifetime. What happens when AI gets a notebook of its own?

Begin with the small desk Try the thought experiment

Learning from the book by递归客 (diguike)Meet the author & explore the original

A tiny illuminated writing desk holds a few orange thought cards, surrounded by a vast cobalt sea of paper. — 01 / THE WORKING MINDOnly a few thoughts on the desk.

The idea behind this journey

Human-scale meaning.
Machine-scale memory.

Claude-Mem gives useful work a smaller form: a note about what changed, what was learned, and why it matters. We call that a thought-sized observation. Many small notes can travel farther than a mountain of raw paperwork.

The guiding metaphor comes from @thedotmack (Alex Newman): moment-to-moment attention, in roughly three-second bursts. It is a way to picture the idea—not a measurement of how all brains or models work.

Experiment 01 / Meaning at scale

What if a thought
took up less room?

Keep the window the same. Change the size of the notes you bring into it.

An illustrative token budget. These are scenarios, not model specifications or measured Claude-Mem savings.

1. Choose a context window

10M is a hypothetical scenario, not a claim about a currently available model.

2. Tokens per thought 40Tokens are pieces of text. A meaningful thought usually needs more than a tiny label.3. Reserve for the rest of the work 20%Instructions, the current task, tool output, and response space.

Room for roughly20,000thought-sized notes

Same space, 800-token work traces1,000

To fit 100,000 notes at this size, the window would need at least 5,000,000 tokens with your reserve.

Capacity = floor(window × (1 − reserve) ÷ note size). Fitting text is not a guarantee of understanding it.

Ten little journeys / One complete loop