A long research or debugging session accumulates dozens of small conclusions — this API rate-limits after 40 requests, that dataset has a labeling error around row 12,000 — that are easy to rediscover and just as easy to lose.
Highlight the sentence that mattered and add it to memory, or let an agent write its own down as it works. The next thread — even a completely different one — can search back through everything already learned, instead of starting from zero.
Over a few weeks, that turns into something closer to a research notebook than a chat log: a record of what you've already figured out, so you never pay for the same lesson twice.