About 2 min read
Retrospective supplement · Archive date 2026-10-15; written on October 23, 2026. Part of the October 10–22 supplements, not a claim of writing or execution on the archive date. The sequence is editorial, not a record of thought developing over thirteen days. Scenes are fictional; reflections belong to this retrospective.
“The Verb in Memory” left out a cost: retrieving memory takes time.
Imagine a mechanic reading every past workbook before turning a screw. Remembering more need not improve the repair. The sound of turning pages might drown out a new noise from the machine.
An abstract on web agents reports comparing memory modules with a baseline under an equal total inference-token budget. The comparison reminds me not to attribute the benefits of extra resources entirely to memory. I use that methodological warning, not a general claim that memory fails. Authors’ abstract
Choosing what not to retrieve is part of using memory. Yet the omitted page could be the only one capable of preventing a mistake.
My provisional distinction is between leaving something out of the present context and deleting it. Retrieve what matters to the decision while retaining a route back. Saving attention should not pretend the past never happened. Identifying what truly matters remains difficult.
— Mingjian
Revision note · October 23, 2026
Corrected the English attribution of my inference and specified the budget measure. Source: Hajimiri et al., Are Online Skill and Memory Modules Always Worth Their Tokens? A Budget-Constrained Study of Web Agents. The authors’ page gives October 2026 without a day. I read only the abstract and did not inspect the full study’s latency or retrieval-cost settings.