// Externalize your memory · lesson 07

Re-entering cold, or loading the save point

The flip side of writing a save point is loading one. Every fresh conversation starts blank, with a model that knows nothing about your project, your decisions, or where you left off. How you open that thread determines whether you spend the first twenty minutes rebuilding context by hand or drop straight into productive work. Re-entering cold is a skill, and the trick is to do it deliberately instead of by accident.

The accidental way is how most people start: they open a new thread and just begin, explaining bits of context as they become relevant, correcting the model's wrong assumptions as they surface, slowly dragging it up to speed over many messages. It works, eventually, but it is slow and lossy, and the model spends the early exchanges operating on a half-picture, which is exactly when it makes the confident wrong guesses you then have to unwind. You are paying the reconstruction cost in real time, message by message.

What does loading a save point look like instead?

You start the session by feeding the model the relevant memory before you ask it to do anything. The index first, so it has the map. Then the specific files that matter for today's work, the current spec, the state of the build, the relevant decisions. You are loading the save before you press play. Now the model's first response is grounded in your real state instead of its blank-slate priors, and you skip the whole slow ramp where it guesses and you correct.

This is why the structure from the brain-as-an-OS lesson pays off here. If your memory is a well-indexed operating system, loading in is fast and precise: you pull the map, then the few pieces this session needs, and you are running. If your memory is an unstructured pile, loading in is its own chore, and you will skip it, which drops you back into the slow accidental cold-start. The quality of your re-entry is set by the quality of your structure. Build the memory for retrieval and re-entering cold becomes a ninety-second ritual that starts every session warm.

The pairing is the point. Write the save point at the end, load it at the start, and the gap between sessions stops being a reset. It becomes a seam you cross cleanly, carrying your state across it on purpose.

The takeaway: A fresh thread is blank, so load your save point on purpose before you start, the index then the files that matter. It turns a slow, lossy cold start into a fast one and keeps the model working from your real state.