MEMORY HIERARCHIES FOR CONTEXT

STRUCTURED NOTE-TAKING Structured note-taking means the agent writes organized notes, outside the context window, as a form of external memory. External memory is memory stored outside the model's window, in a place the harness can read back later. Why notes instead of relying only on summaries? An automatic summary is written by a process that does not always know what will matter later. Notes written by the agent as it works, in a fixed structure, tend to capture the right things, and they stay exact because they are updated, not repeatedly re-paraphrased. A good notes file has predictable sections. For example, a file might contain a list of completed steps, a list of open questions, a list of decisions with reasons, a list of important facts or file locations, and a single line stating the next action. Because the format is consistent, both the agent and the harness can read it quickly. A very common version is a plain text or markup file saved in the working folder, with a name such as progress notes, that the agent updates after every major step. If the context is ever compacted or the whole process is interrupted and restarted, the first action on resuming is to re-read the notes file. The agent then knows exactly where it left off. This is much more reliable than hoping a summary captured the status. Structured notes also support long jobs that span many separate sessions, even days. Each new session begins with no memory at all, but by reading the notes it can pick up the thread, in the way a night-shift nurse reads the handoff chart left by the day shift. When designing a notes format, follow these principles. Keep it short enough to re-read in full at the start of every session. Make each entry self-contained, so it makes sense without surrounding context. Record decisions together with their reasons. Mark status clearly, such as done, in progress, or blocked. Put the next action at a fixed place so it is easy to find. And have the harness validate the file's structure, so a malformed update does not silently corrupt the agent's memory.