Working memory

Working memory is the context window itself. It is ephemeral, high-bandwidth, and limited. Everything the agent is actively reasoning over lives here, briefly.

Its characteristic failure mode is attentional dilution, also called the "lost in the middle" effect: when the window gets too crowded, relevant content gets ignored even though it is technically present. The "Practical Guide" notes this is hard to debug precisely because the model "has" the information but is not using it, which is the same phenomenon as context rot from the context-engineering material.

The practical workaround that practitioners (and the author) reach for is mundane but effective: start a new thread for a new chunk of work. The advice is vivid: you do not keep a coding assistant open all day across 20 different tasks, because its working memory degrades and its performance drops. Fresh context beats a stuffed one.