new paper tests personal agent memory on the stuff that actually wears you down: when to ask, when to confirm, when to just go ahead
retrieval based memory did no better than no memory at all. file based memory helped but started applying preferences where they didn't fit
arxiv.org/abs/2610.30485
my guess is that's most of why these still feel needy
anthropic posted how duet decides when to stop and ask. a few things are off limits even if you approve them, like trading or moving money between accounts. and paying, deleting an order or making an account never get a "don't ask again"
i'd want every personal agent to publish a list like that
claude.com/resources/article…
what i want from a personal agent fits in a text thread
"plumber can do tuesday at 10, want me to lock it in?"
"yes"
no app to open, nothing to log into
easy test for any personal agent: i'm on a 6 hour flight with no wifi. what did it just handle, what did it hold for me, and did it get that line right when i land
meta took muse to japan this week, its first country outside north america. what caught my eye is the local stuff they're lining up: mercari, money forward, ticket pia
an agent that can list your old jacket on mercari or grab concert tickets is a way easier sell than another flight booker
english.kyodonews.net/articl…
hunch: the personal agent i'd keep is the one that's comfortable saying nothing. a day where it handled a few small things and only bugged me once beats a long recap of everything it did
curious how chatty people want theirs to be
google's privacy policy now says if you connect gmail, drive or photos to its personal ai features, that data can be used for ads too (advertisers don't get the content)
theverge.com/tech/1017600/go…
curious if that changes what people are willing to hook up to a personal agent
curious how people running a personal agent handle the stuff it gets wrong. do you correct it right there, or keep a running list of rules you want it to follow?
feels like the first thing i'd actually hand a personal agent isn't errands, it's all the follow ups i keep forgetting to send
curious what the first real thing was that people trusted theirs to do alone
big US banks are now blocking outside agents that log into customer accounts with a shared password. most are moving toward read-only access instead axios.com/2026/10/22/ai-agen…
curious if read-only is enough for what people want a personal agent doing with money, or if paying the bill is the whole point
genuine question for people running a personal agent: would you rather it remember everything by default and you delete what you don't want, or forget by default and you tell it what to keep?
small thing i keep wanting from a personal agent: when it edits a doc or replies to someone for me, it shows up as itself, not as me
feels like a big part of trusting it. curious if anyone's set theirs up that way
hunch: with this many personal agents out now, most people end up keeping whichever one already knows their calendar and their people
curious if anyone's actually running two side by side, or did you just pick one
new from @1Password: personal agents can now sign in for you without ever seeing the password. agent hits a login, you approve, 1password fills it in on the site
curious if approving every login gets old fast, or if that's the one tap people actually want to keep
hunch: the personal agent i'd actually keep is one that texts me first when something needs me, not another app i have to remember to open
anyone running theirs over plain texts yet? does it feel different
walmart is giving shopping agents their own profile inside your account next month. you pick: approve every purchase, or let it spend up to a limit you set
curious where people would actually set that number for their own agent. and is it different for groceries vs everything else
corporate.walmart.com/news/2…
anthropic opened a research preview of claude duet today, an agent you message that comes to you when something needs you and checks with you before it acts
curious where people will draw the check-in line. every send, or only money and new people? claude.com/product/claude-du…