Writing a scratchpad is the easy part. A store that carries an agent’s state across sessions, whether that is a file in a coding agent’s workspace or a record in the datastore behind a deployed system, can be stood up in an afternoon, and on day one it does exactly what it promises. The difficulty is not authorship. It is keeping…

An agent has no memory of its own work. Each session opens with a blank context window, and everything the previous session learned, the approaches it tried and abandoned, the naming it settled on, the reasons behind the choices already made, is gone the moment that session ends. This is not a defect to be patched. It is the shape of how these…

Once you accept that a tool’s raw output should not enter context unchanged, the reflex is to treat trimming as a single technique with one implementation: strip the payload down to what the agent needs and pass along the remainder. That framing hides the actual decision. There is no single trim. There are three distinct strategies, they trade away different resources…

The most-watched numbers in an agent system are usually the prompt and the model’s own output. The instructions get versioned, the responses get graded, the system prompt gets audited line by line. Meanwhile the largest and least-examined source of tokens in a long-running agent is often the raw output of the tools it calls. A single API response can be…

Once you accept that a summarizer will drop the exact values a long session depends on, the fix is easy to describe and easy to half-build. Pull the facts out of each turn before compression runs, write them into a durable region the summarizer never touches, and the values survive while the narrative is free to shrink. Stated that way, a…

Long conversations outgrow the window that holds them. A session that runs for hours, an agent that works a task across dozens of turns, a support thread that keeps accumulating, all of them eventually carry more history than the model can take in at once. The standard answer is to compress the older part of that history into a shorter…

There is a comfortable assumption behind a lot of context handling: the window has a ceiling, so anything up to that ceiling is fair game. Retrieve a few extra passages in case they help. Keep the whole conversation because trimming is work. Carry a generous system prompt because instructions feel safer when there are more of them. The window fits it…

A language model has no memory of you. It has no memory of the last thing it told you, no running notion of the task, no internal notebook that persists from one call to the next. What looks like memory is an illusion reconstructed on every request. The single place that illusion is assembled is the context window, and that window…