The Quiet Shift: Three Assistants, Two Years, Memory On by Default
Between early 2024 and March 2026, the three major AI assistants made the same journey: from stateless chat windows to persistent personal memory that is on unless you turn it off. The shift happened gradually enough that most people never marked the moment their assistant started keeping a file on them.
The progression made sense at every step. Memory began as a convenience, a way to stop repeating your name, your stack, your dietary restrictions, and each expansion was easy to welcome because each expansion made the assistant better. An assistant that remembers your project does not need it re-explained. An assistant that remembers your writing voice drafts closer to what you would have written. By the time all three vendors had made memory the default rather than the opt-in, the feature had stopped being a novelty and become part of what an assistant simply is. Users did not resist this, and there was little reason to: statelessness was never a virtue, only a limitation.
It is worth being clear-eyed about why the vendors converged here so quickly, because the answer is not sinister and not entirely innocent either. Models themselves are becoming harder to differentiate; what one lab ships, the others match within months. Accumulated context is different. Your year of preferences, projects, and corrections exists in exactly one place, and it cannot be matched by a competitor shipping a better model, because the competitor's better model does not know you. Memory is the rare feature that improves the product and deepens the moat with the same lines of code, which is precisely why all three raced to build it, and why the race matters to you even if you never think about vendor strategy at all.
The result is that many people now carry a substantial, largely invisible asset inside their assistant of choice: a compounding record of their working life, built one conversation at a time. The question this article is really about is not whether that asset is valuable. It clearly is. The question is who holds it, on what terms, and what happens when you want it back.
Three Memory Designs, and What Each Gets Right
Underneath the similar marketing, the architectures split into two genuinely different philosophies, and each one is a defensible answer to a real design question: should memory be an invisible sense, or a legible record?
OpenAI and Google chose the invisible sense. ChatGPT and Gemini store what they learn about you in opaque, vector-backed form, a representation built for the model's retrieval rather than for your reading. You can see fragments, short summaries the system chooses to surface, and you can delete or disable the whole apparatus, but there is no document you can open that shows, in full, what the assistant knows about you. The logic of this design deserves credit. Memory that works like human familiarity, ambient, unspoken, never presented for review, produces the smoothest possible experience. Nothing to manage, nothing to file, no settings page you are obliged to visit. The assistant just seems to know you, the way a longtime colleague does, and for the millions of people who want exactly that, the design delivers it.
Anthropic chose the legible record. Claude stores its memory as human-readable files that you can open, read in plain language, and edit directly, correcting what is wrong and deleting what should not be there. When Claude's memory reached all users, including the free plan, in March 2026, this design went from a differentiator for the few to a baseline available to anyone. Its logic also deserves credit, and it is a different logic: memory that describes you is consequential, so you should be able to audit it the way you can audit anything else written about you. The cost is a little visibility of the machinery. A file you can read is a file you know exists, and some people would rather not think about the mechanism at all.
This is the honest trade, and it is worth stating without a thumb on the scale: transparency versus frictionlessness. The opaque design is smoother precisely because you never see it, and the readable design is more accountable precisely because you do. Neither vendor is wrong about what its users want, and if the story ended at user experience, you could pick by temperament and be done. The story does not end there, because memory is not only an experience. It is also a record, held by someone, and records raise questions that user experience does not answer.
The Ownership Questions: Read, Correct, Export, Redirect
Strip away the interfaces and the branding, and what separates the memory systems is four questions. They sound simple, and they are the whole game.
- Can you read it? Not a summary of it, the thing itself. On this question the designs split cleanly: Claude's memory files are readable in full, while what ChatGPT and Gemini hold about you is visible only in the fragments each product chooses to show. A record you cannot read is a record you are taking on faith.
- Can you correct it? Assistants infer, and inference gets things wrong: a stale fact, an abandoned project treated as current, a throwaway remark promoted to a preference. Editable memory lets you fix the record directly. Opaque memory offers deletion and hope, remove entries, or reset entirely, and trust that the correction took.
- Can you export it? Not delete it, take it. A copy in a form that is useful outside the product that made it. For most of the memory era the answer everywhere was effectively no, and as of this writing it has improved to partially, which is what the March 2026 tools are about.
- Can you point another tool at it? The deepest question of the four. Your context is only infrastructure if any tool you choose can use it. Memory that answers only to its own assistant is not your infrastructure, it is the vendor's, however pleasant it is to use.
March 2026 brought a moment that looked, briefly, like the answer to the last two questions. Within a few weeks of each other, all three vendors shipped memory import or portability tools, with Claude's Import Memory able to pull context across from ChatGPT and Gemini. The near-simultaneous timing says something on its own: portability had become a competitive necessity, a door each vendor opened mostly so that users could walk in through it, with the exit a secondary concern. Still, the tools are real, they work, and switching assistants in 2026 is materially less costly than it was in 2025. That is progress worth acknowledging plainly.
It is also less than it appears. What moves in these transfers is a summary, a condensed digest of what the source assistant held, not the full accumulated record. Commentators reached for the same image independently: it is like getting a photocopy of one page of your notes. The new assistant learns roughly who you are. It does not inherit the year of specifics, corrections, and accumulated nuance that made the old one feel like it knew you, and nothing currently on offer does. Portability, as of this writing, moves the gist and leaves the substance behind.
Lock-In, Stated Fairly
None of this requires a villain. Memory that works only inside its own assistant is not a bug or an oversight. It is the design, because memory is a retention feature, and it is possible to say so without accusing anyone of bad faith.
Every vendor building memory is doing two things at once, and both are real. They are making their assistant better, measurably, in ways users feel every day. And they are raising the cost of leaving, because every month of accumulated context widens the gap between the assistant that knows you and the identical model that does not. These are not two faces of a deception, they are one feature seen from two sides, and the vendors are doing what any rational company with a retention lever does. The point of naming the dynamic is not to condemn it. It is to make sure you price it in.
So price it in concretely. Suppose a year from now a competitor ships an assistant that is meaningfully better for your work, better reasoning, better tools, better price. The model is available to you the day it launches. Your context is not. The new assistant does not know your projects, your preferences, how you like disagreement handled, which explanations landed, what you decided last spring and why. An import tool hands it the one-page summary, and then you spend weeks or months rebuilding through repetition what the old assistant had accumulated, or you stay where you are, with the slightly worse assistant that knows you, and most people, most of the time, will stay. That is lock-in operating exactly as intended, through genuine value rather than contractual force, which is what makes it both legitimate and worth defending yourself against.
The Neutral-Home Pattern: Your Record, Their Caches
There is a way to hold all of this that does not depend on predicting which vendor wins, and it borrows the oldest idea in computing: keep one canonical copy, and treat everything else as a cache.
The pattern is simple to state. The permanent record of your thinking, your notes, decisions, research, references, the outputs of your AI conversations that were actually worth keeping, lives in a library you own, in a neutral home that no assistant vendor controls. Every assistant's memory is then demoted, gently, to what it actually is: a convenience layer, a cache of your context inside one product, rebuildable at any time from the real record. Caches are wonderful. You never protect a cache, never treat it as the only copy of anything, and never grieve when one is invalidated. If your canonical record is yours, then switching assistants becomes a question of which tool is better this year, not a question of how much accumulated self you are willing to abandon.
The pattern flows in both directions. Downstream, when a conversation with any assistant produces something that matters, a decision, a plan, a draft that works, an explanation that finally made something click, that conclusion gets saved into your library at the moment it appears, because an assistant's chat history and its memory are the two places most likely to quietly become the only home of your best thinking. Upstream, when you need an AI to know your context, you hand it the relevant material from your library, this document, this note, this set of decisions, rather than relying on whatever the assistant's memory happened to retain. Your record becomes the source of truth that any AI can be given, today's or a better one next year, and no assistant's memory is ever the only place something lives.
What the neutral home requires of you is honesty about one thing: it is a habit, not a product feature, and habits survive only when they are nearly effortless. If saving the conclusion of a conversation takes six steps, you will stop doing it within a week, and the assistants' memory will quietly resume its role as the only record. The pattern stands or falls on the cost of the save.
Where Mindly Fits
Mindly is a native macOS app built to be that neutral home: a local, vendor-neutral library for the record that should not live inside any assistant.
The save costs one shortcut. Press ⌘M and whatever is in front of you goes into your library, a note, a link, a PDF, a file, a screenshot, a voice memo, which makes capturing the conclusion of an AI conversation a two-second act instead of a filing chore. The librarian work that memory systems do invisibly happens here in the open: AI tags each item by topic, summarizes long content, transcribes audio, reads the text inside images, and links related items together, and every one of those tags, summaries, and connections sits on an item you can open, edit, or delete. When you need something back, you search in plain language and the search works by meaning, so the decision you remember as the one about switching costs surfaces even when the words do not match. You can open any saved item and chat with it, asking questions answered from your own material rather than from a model's general knowledge, and a mind map shows how the library connects.
The ownership questions from earlier in this article have direct answers here. Can you read it: everything, it is your library. Can you correct it: every item and every tag. Can you export it: yes, to standard formats, because a canonical record you cannot take with you would be the same trap with better branding. Can you point another tool at it: that is the entire point. The library lives on your Mac, AI processing runs over encrypted channels and your content is not retained after the request completes, and no account is needed to start. The free tier holds 25 items, and Mindly Pro removes the limit at €7.99 a month or €44.99 a year. The assistants will keep fighting over whose memory knows you best, and you can let them, comfortably, because the record they are all approximating will be sitting in a folder you own.
Free for macOS, no account needed. Give the record of your thinking a home that no assistant owns. Download Mindly →