Main idea

A session memory writes itself: the hook runs after every exchange, without anyone thinking about it. It doesn't reread itself for all that — you have to run /memory_load at the start of a session for the agent to know where the work stood. The asymmetry looks like a limitation nobody had time to lift; it is a choice.

The reason is that each gesture has a cost. Writing to a file bothers nobody: at worst, you recorded something useless you will never read. Loading, by contrast, takes up room — a project's memory injected into a session opened for a three-line question clutters the context with unrelated material, and the assistant answers less well for having received it.

The two movements therefore don't follow the same rule. What has no side effect can be automatic; what consumes a shared resource requires an intention. It is also what lets you pick work back up without having to endure its memory every time you open the tool.

Why it matters

This overturns an apparent obviousness — that a complete memory system would be one that reloads itself — by showing that automatic loading would degrade precisely the most frequent uses, the short ones.

And it gives a general criterion for deciding what to automate: not what you repeat, but what costs nothing to whoever is on the receiving end of the repetition.

Nuances and limits

A manual gesture gets forgotten. A memory correctly written but never reloaded is of no use, and nothing flags that you have just reworked blind on a subject where everything had been recorded.

And the asymmetry assumes the human knows when they need it — which is true on a Monday morning about a bug left hanging, less true about a project picked back up after three months.

Open questions

• What do you offer someone who doesn't know a memory exists about what they are doing?