ai memory is everywhere in 2026. none of it is yours.

2026 was the year ai memory stopped being a feature and became table stakes. openai shipped a background memory system called dreaming on june 4. anthropic replaced claude memory's rolling summary with editable entries on july 10. google rebranded gemini's personal context as gemini personal intelligence in january. microsoft finished rolling copilot memory across m365 by may.

five products, one year, the same idea. every one of them promises an ai that remembers you.

and a question almost nobody is asking: whose memory is it?

what each assistant actually remembers

all five make the identical promise, which is that you stop repeating yourself to a machine. the implementations differ in ways that only start to matter once you try to leave.

chatgpt memory is the most automatic of the five and the most closed. the rollout of dreaming replaced the old list of manually saved facts with a process that reviews your chats after they end and updates what the assistant knows, without you ever typing "remember this."

anthropic went the other way. claude memory is a list of individual entries you can read, correct or delete one at a time, and it now spans both chat and longer working sessions. anthropic's own announcement frames it as removing the need to re explain context.

google made the boldest portability move. in march 2026 gemini added import tools for chatgpt and claude history, letting you carry your context in.

the portability trap

memory flows toward whoever is behind. that is the whole pattern.

claude imports from chatgpt, gemini and grok. gemini imports from chatgpt and claude. chatgpt, the market leader, imports from nobody. so the direction of travel is out of the leader and into the challengers, and it is a one way trip. whatever you build up in claude after the import does not go back.

this is not a bug in anyone's product. it is what happens when memory lives on the vendor's side.

the practical test is whether you can export ai memory on demand, in a format something else can read, without asking a competitor to scrape it for you. on that test all five score badly.

why provider memory is thinner than it sounds

every launch post skips the same thing. even a perfect assistant memory holds one slice of your working life, which is the part you typed into that particular box.

it does not contain the pdf you skimmed and closed. the figure on someone's slide. the decision made out loud on a call. the error message you googled at 11pm and never mentioned again. per the anthropic economic index, the single most common work task in ai conversations globally is "search electronic sources, such as databases or repositories, or manual sources for information," at 4.95% of all sampled conversations.

work accounts for 43.36% of classified conversations, personal life for 40.2% and coursework for 16.45%. of all usage, 51.38% is augmentation rather than automation, meaning long back and forth sessions rather than one shot answers.

the topic mix tells you what is missing. content creation is 22.72% of conversations and software development 11.51%, but knowledge retrieval and enterprise search is 3.61% and personal ai assistant is 2.86%. the assistant everyone is building memory for is 1 in 35 conversations, because it has nothing of yours to work from.

so the typical session is a person thinking out loud against someone else's corpus. never against their own week.

ask any of the five what you were working on last tuesday afternoon. they will tell you what you told them, which is a very different thing. we wrote about that gap in what is an ai memory app.

who can read it, and for how long

provider memory sits on provider infrastructure under provider policy. three consequences follow, and none of them are hypothetical.

policy changes. anthropic began leaving categories out of claude's memory by default on august 25, 2026, across 6 categories: health conditions, race, ethnicity, religious beliefs, political views and gender identity.

that is a good change. it also confirms those 6 things were in scope before somebody decided otherwise.

ownership changes. rewind spent two years telling mac users their screen history was private and theirs. the company renamed itself limitless, and in december 2025 meta acquired it and killed the mac app.

capture stopped on december 19, 2025. users in 7 countries including the eu, uk and south korea lost access immediately, with 14 days to export before deletion.

products end. copilot memory is scoped to a tenant your employer controls. when you leave the job, you leave the memory.

what a memory you own looks like instead

putting the archive on your side of the line changes all three problems at once, and it does not mean giving up memory.

that inverts the three problems above. the record is of your work rather than your chats, so it covers the calls and documents and tabs. it does not belong to a vendor, so an acquisition cannot switch it off. and it is not locked to one assistant, so switching from chatgpt to claude costs you nothing.

remynd takes that approach on a mac. it captures the focused window, runs ocr on it locally, and keeps the index on your machine, so the archive is yours whether or not you are paying anyone.

an agent can then query it. we shipped that path in august 2026, and claude code and codex can now ask your own history a question through a local connector.

be precise about the privacy claim, because plenty of tools in this space are not. remynd keeps capture, ocr and storage on your mac by default, and you can exclude specific apps or sites entirely. it is not a claim that nothing ever leaves the machine, since sign in and the cloud agent do reach the network. the honest version is that the archive lives with you. that distinction is spelled out on our security page and in private ai on your mac.

how to decide, in four questions

1. can you read every entry? if you cannot enumerate what it knows, you cannot correct it.
2. can you delete one thing without deleting everything? claude's july change exists because the old answer was no.
3. can you take it with you? check the import list in both directions, not just the one on the marketing page.
4. does it cover your work or only your chats? this is the big one, and for all five products above the answer is chats.

provider memory is genuinely useful and worth turning on. just do not mistake it for a record of your work. it is a record of one conversation you happened to have in one company's product, held on their terms, for as long as they feel like holding it.

download remynd for mac and keep the other copy yourself.