Nice summary of techniques for effective memory design for agents. I’ve been doing a fair bit of building in this space now and what is very noticeable is that building these features is trivial but tuning and carefully evaluating them is hard, especially while minimising spend and delivering an unobtrusive product experience (neither if which are really mentioned here). There’s a noticeable difference between what’s being published by the labs and the research papers in this space - the research papers lag the frontier significantly and the behaviour is very different between model generations (see Anthropic’s own notes here comparing 4- series models and 5-) which often means the research is not a reliable guide on design. I think it is very easy to over engineer on that basis when I suspect very simple things actually deliver the right performance/cost/UX trade off. Interesting space.