claude memory or files: where your business context should live.
on august 25 claude's memory stopped being a chat-only thing and started following you into cowork. that changes what is worth putting in it, and what still belongs in a file you control.
claude's memory and your context files do different jobs. memory holds the facts about you that should follow you into every conversation, and since august 25, 2026 it follows you into cowork too. files hold the things that must be exact, versioned, and readable by other people. put your preferences in memory. put the record in files.
most people will read the update as "claude remembers more now" and change nothing. the useful read is narrower: one category of context just got easier to carry, and it is not the category that keeps costing you money when it is wrong.
what changed on august 25, 2026
three things changed, and the headline one is the least interesting. memory now works across claude chat and claude cowork rather than living separately in each, so context you built up in chat carries into an agent session. every stored item is listed topic by topic under settings, memory, where you can read, edit or delete it individually. and the capture moved earlier: instead of summarising a conversation once it ends, claude writes topics while you are still talking.
"claude now adds topics to memory as you chat" (Anthropic, august 25, 2026)
memory is on by default on the free, pro and max plans across web, desktop and mobile, while team and enterprise administrators decide whether it is available at all (Anthropic, august 25, 2026). that plan split matters more than it looks: if you run a business on a team seat, your memory behaviour is an admin setting, not a personal one, which is one more reason not to make it your system of record.
memory or files: the side by side
memory is a convenience layer and files are the record. both are context. only one of them is something you can check, hand to a contractor, or defend in a room.
| what you are deciding | claude memory (topics) | a context file |
|---|---|---|
| what it holds best | standing preferences and facts about how you work | anything that has to be exact: prices, offer specs, client details |
| who can read it | you, on your account | you, a contractor, another tool, and future you |
| how it gets in | claude adds topics while you chat | you write it once, deliberately |
| how you fix it | edit or delete the topic in settings | edit the file and save |
| when it is wrong | you may not notice until it shows up in output | the file is the thing you check the output against |
| survives a tool switch | no, it stays with the account | yes, it is a file |
the six-file version of the right-hand column is in files not chats: the 6-file context system, and the knowledge-base build is in how to build a claude project knowledge base that will not lie to you.
the four-question test for what goes where
run any piece of context through these four in order, and the first yes decides it. the test is deliberately biased toward files, because the failure mode of a file is that you forget to attach it, and the failure mode of memory is a wrong fact you never see.
- would getting this wrong cost money or credibility? file. prices, contract terms, client names and offer specifications need to be checkable, not recalled.
- does anyone else need to read it? file. memory is tied to your account and cannot be handed to a contractor, a teammate, or a different tool.
- does it change more than once a month? file. fast-moving detail is safer in one place you edit on purpose than spread across topics that update while you talk.
- would you otherwise retype it in every conversation? memory, if it survived the first three. how you like things written, what you are working on this quarter, which tools you actually use. this is what memory is for and it is genuinely worth setting up.
the layered version of the memory side, written before this update, still holds: how to set up claude's memory. and the wider discipline of deciding what claude gets to see at all is in context engineering for operators.
the part worth being careful about
the risk is not that claude stores too much, it is that memory now reaches the surface that acts. chat suggests. cowork does multi-step work. a wrong fact sitting in a memory topic used to produce a wrong sentence you could catch on the way past. now it can seed a task. that is not an argument against the feature, it is an argument for reading your topics list once before you point cowork at anything real.
on the privacy side, anthropic excludes sensitive topics by default, naming health, race, ethnicity, religious beliefs, politics and gender identity, and you have to switch on "include sensitive topics in memory" to change that (Anthropic, august 25, 2026). some categories are never stored even with that setting on: government identification numbers, criminal history, immigration status, and anything against the acceptable use policy. for business use that boundary is a helpful hint about intent. memory is built for how you work, not for your records.
the surface-by-surface version of this decision, written when chrome gained full cowork sessions, is in which claude surface for which task, and the safety rules for letting cowork run are in how to use claude cowork safely.
common questions
should i put my business context in claude's memory or keep it in files?
both, split by the four-question test above. preferences and standing facts in memory. anything exact, shared or versioned in files. if a wrong detail would cost you money or credibility, it belongs somewhere you can check rather than somewhere you hope is still accurate.
what actually changed with claude memory on august 25, 2026?
three things. memory now works across claude chat and claude cowork rather than being separate to each, so context built in chat carries into an agent session. everything claude remembers is listed topic by topic under settings, memory, where you can read, edit or delete any individual item. and claude adds topics while you chat instead of summarising a conversation after it ends. memory is on by default for free, pro and max plans, while team and enterprise administrators control availability.
does claude store sensitive information in memory?
not by default. sensitive topics including health, race, ethnicity, religious beliefs, politics and gender identity are excluded unless you switch them on. government identification numbers, criminal history and immigration status are never stored either way.
how do i stop claude remembering something wrong about my business?
settings, then memory, find the topic, edit or delete it. the correction carries forward instead of needing to be repeated. if the same wrong fact keeps returning, that is a signal it should have been a file.
does this replace project knowledge or a claude.md file?
no. memory is about you across projects. project knowledge is about one body of work. they answer different questions and the august 25 update did not merge them.
the takeaway
about 85% of people have access to AI and about 25% use it on real work (IBM 2026 Global CEO Study, cited in the ai adoption gap). that gap is not closed by a model remembering your name. it closes when the context a business actually runs on lives somewhere you can read, correct and hand over.
so treat august 25 as an invitation to do one boring thing: open settings, memory, and read the topics list end to end. delete what is stale, correct what is wrong, and move anything exact into a file. twenty minutes, once. then memory holds your preferences, files hold your record, and the two stop fighting.
want the setup, not the trial and error?
the ai builder toolkit is the set of claude skills that build the files side of this: the context system, the knowledge base, the checks that catch a wrong fact before it ships. ready to install and run in claude today.
see the toolkitbuilding this alongside other people works better than building it alone. the ai builders lounge is where the weekly builds happen.
or just follow along. new field notes most weeks on x, instagram, and tiktok.