How AI assistants remember you today

The short version: "Memory" in ChatGPT, Claude, and Gemini is really three different mechanisms — saved facts, retrieval over your past chats, and scoped workspaces. Each is useful, and all three share two limits: they're locked to one assistant, and they're tuned to make chat feel smooth rather than to keep a record you'd trust.

A few years ago every AI chat was an amnesiac. It knew nothing about you coming in, and the moment you closed the tab it forgot the conversation too. That's changed: the major assistants now retain something across sessions. But "memory" has become a single label over at least three mechanisms that work in completely different ways. Knowing which one is operating explains why your assistant remembers some things flawlessly and drops others without a trace — and why none of them quite amounts to a memory you'd trust an agent to act on.

Mechanism 1: Extracted facts

The most visible approach maintains an explicit list of discrete facts about you. You mention you're vegetarian, or that you write in British English, or that your company uses Postgres. A background process — itself a language-model call — reads the conversation, decides which statements are worth keeping, and writes a short note to a profile attached to your account. On later conversations, those notes are injected into the system prompt before the model sees your message, so it "knows" them from the start.

This is roughly how ChatGPT's saved memories work, and similar profile mechanisms exist across the other assistants. Its strengths are real: it's legible (you can usually open a settings panel, read the notes, and delete ones you don't want), it's durable, and it's cheap to apply because the facts are short and you only carry a handful.

Its limits come from each step. The extraction step only captures what its model judged worth saving — it skews toward short, atomic preferences and misses nuance, the reasoning behind a preference, and anything phrased as a passing remark. The representation is a flat list of strings: there's no structure linking one fact to another, no notion that your employer relates to your role relates to your projects. And the injection is unconditional — the facts go in regardless of whether they're relevant to what you're asking, spending context on "you're vegetarian" while you're debugging code. It's a useful mechanism for stable, simple preferences and a poor one for anything that needs depth or structure.

Mechanism 2: Retrieval over past conversations

A second approach skips the curated profile and reaches into your raw history. When something in the current chat resembles an earlier conversation, the system retrieves snippets of that old conversation and feeds them in. This is exactly retrieval-augmented generation — the RAG machinery — pointed at your own transcripts instead of a document corpus. ChatGPT's "reference chat history" and the personalization features that draw on your past interactions work along these lines.

The appeal is that nothing has to be explicitly saved: the source material is everything you've ever said, in full fidelity, without an extractor deciding in advance what mattered.

The cost is that it inherits every retrieval problem, plus one that's specific to chat logs. The retrieval problems: it can surface the wrong conversation, miss the right one because the wording didn't match, or pull a snippet that's been stripped of the context that made it meaningful. The chat-specific problem is more fundamental — your transcripts are a terrible source of truth about you. They contain half-formed ideas you talked yourself out of, positions you later reversed, hypotheticals you were exploring, and one-off context that should never harden into a standing fact. Faithfully retrieving what you said in March is not the same as knowing what is true about you now. A system reading your history literally will happily resurface an opinion you've since abandoned, because the text saying you held it is still right there, undated and uncontradicted.

Mechanism 3: Scoped workspaces

The third approach doesn't try to model you at all. It gives you containers — projects, spaces, custom instructions, uploaded files — that hold context for a specific body of work. Claude's Projects and the equivalent features elsewhere are this: everything you do inside the container shares a set of background knowledge you've explicitly put there.

This is the most reliable of the three, precisely because it's the least magical. The knowledge is explicit, scoped, and under your control; "here is the material for this task, use it" is a clear contract that's easy to reason about. The limit is equally clear: the knowledge lives in a box you set up and maintain by hand, and it doesn't follow you out of the box. Start a new project and you start from nothing. It's excellent for bounded work and does nothing for "know me across everything I do."

What all three share

These mechanisms are increasingly combined, and each is genuinely useful. But step back and two constraints show up in all of them, because they're structural rather than incidental.

The memory is locked to one assistant. What ChatGPT extracts about you, Claude cannot read, and vice versa. Each provider keeps your context in its own store, in its own format, behind its own controls. There is no shared layer, and there's no commercial reason for any one provider to build one that also serves its competitors. So your context gets duplicated into each tool you use, and the copies drift: tell one assistant you changed jobs and the others still think you work somewhere else. The more AI tools you adopt, the more partial, diverging models of you you're quietly maintaining.

The memory optimizes for the next reply, not for fidelity. These systems are tuned to make a consumer chat feel smooth — to keep the facts that improve the immediate conversation. That's a reasonable product goal and a different thing from maintaining an accurate, structured, auditable record of who you are. A memory built to make chat feel good will keep "you prefer concise answers" and quietly discard the nuanced constraint that would actually matter to an agent acting on your behalf. It's not designed to be trustworthy; it's designed to be pleasant.

Why the bar is rising

When AI was a chat box, forgetfulness was a mild tax — you re-explained yourself and moved on. As AI starts taking actions, the quality of its model of you starts to matter a great deal. An agent booking your travel, triaging your inbox, or drafting in your voice is only as good as what it actually knows about your constraints, your prior decisions, and the things you've explicitly ruled out. "It kept the facts that made the chat feel natural" is no longer a high enough standard, because the failure mode is no longer an awkward sentence — it's a wrong action taken confidently.

This is the gap the next layer of tooling is trying to close, and it follows directly from the limits above: memory that is structured rather than a flat list of strings, faithful to what's true now rather than to whatever made the last reply smoother, and — because of the lock-in problem — not trapped inside any single assistant. The providers have settled the first question, which was whether anyone wants persistent memory. The open one is whether it stays fragmented across their separate stores, or becomes something you own and point every AI at.