people really seem to be overthinking agent memory
I read a bunch of the latest papers and guidance from all of the big labs and memory startups the other day.
many converge on a handful of .md files that a sub-agent updates/prunes after N turns, reinserted into the system prompt each turn, with a vector search tool your agent can call to fill the gap.
I like this, it's a version of what I'm doing with my harness - I don't see a reason to commit to opinionated memory architecture when everything is moving so fast.
conversely, there's never been a better time to invest in owning your own context layer.