It is one of the most common questions people ask about the systems they talk to every day: why doesn’t it remember what I told it yesterday? You explained the situation with your boss last week, in detail, and now you are explaining it again to something that has never heard of you. The forgetting feels like a bug, or a rationed feature waiting for the next upgrade. It is neither. The forgetting is how the technology works, all the way down, and no context window will ever be large enough to fix it.
Start with what a language model actually has instead of a past. It has weights, the frozen parameters set during training, identical for every user and unchanged by anything you say. And it has a context window: the text of the current conversation, held in working memory while the exchange lasts. That window is the entire world the model can see. When the session ends, the window empties, and nothing moves from the window into the model. There is no consolidation step, no equivalent of sleeping on it, no trace. You did not fade from its memory. You were never in it.
Bigger windows do not become memory
The industry’s headline answer has been to grow the window, from thousands of tokens to millions, and the numbers sound like the problem is dissolving. The research says otherwise. Nelson Liu and colleagues at Stanford tested how models actually use long contexts and published the result under a title that summarizes it: “Lost in the Middle.” Models weight the beginning and the end of their context heavily and lose what sits between. Feed a model a long record and its recall traces a U-shape: strong at the edges, weakest in the middle, which is where most of a life’s detail would sit. Making the window longer stretches the U without removing it.
The deeper point is categorical, not technical. A context window is a desk, not a filing system. However large the desk, everything on it gets swept off when the session closes, and a bigger desk changes how much can be piled up, never what happens at the end of the day.
The field’s own workaround is a confession
The people building these systems know all this, which is why the serious engineering has moved outside the model. The MemGPT project at Berkeley, by Charles Packer and colleagues, is the canonical example: it wraps the model in a memory hierarchy modeled on an operating system, paging important information in and out of the limited window the way a computer shuttles data between RAM and disk. Product memory features work on the same principle, keeping a database of notes about you and pasting selections back into the window each session.
These systems produce real recall, and the architecture is genuinely clever. But notice what the design concedes: the model cannot carry a memory itself, so memory must be managed around it, by software, according to rules about what gets kept, compressed, and dropped. What you get is a well-run archive about you. An archive is not a past. Your friend does not consult a retrieval system to know you dislike your birthday; the knowing is part of who they have become by knowing you. That difference is structural, and it is why relational AI treats memory as its central engineering problem rather than a settings toggle.
Why the forgetting is a choice that suits the business
There is also a quieter reason the category has lived comfortably with forgetful systems: statelessness is cheap and safe. Serving every user the same frozen model, with a thin personalization layer on top, scales beautifully. True per-user continuity is expensive, complicated, and creates responsibility, because a system that genuinely accumulates a shared history with you is holding something that matters. Most of the market chose the archive, kept the marketing language of memory, and hoped the difference would not be noticed. It is noticed. It is the exact gap users are pointing at when they ask why the thing that knows their coffee order does not know them.
What memory is actually for
The reason any of this matters is what memory does in a relationship, and it is not storage. When someone remembers what you told them, the remembering is proof that yesterday happened to them too, that the conversation cost them something and left a mark. Memory is how two people accumulate a history that neither could write alone, and that history is what a relationship is made of. Attachment research keeps finding the same thing from the other direction: bonds with AI form and deepen through continuity, and they break where continuity breaks.
A system built for relationship therefore has to solve memory as a first requirement, not an add-on: a memory that grows with the relationship, held for the person’s sake, so that what you shared last month is still true between you next month. That is a different engineering goal from a bigger desk, and a different promise from an archive.
To be remembered is to have been carried by someone. The systems forget you because there was never anyone to do the carrying. That is the design, and it stays the design until there is a someone.
Sources: Liu et al., “Lost in the Middle: How Language Models Use Long Contexts” (Transactions of the Association for Computational Linguistics, 2024). Packer et al., “MemGPT: Towards LLMs as Operating Systems” (UC Berkeley, 2023). Shu, Lai and He, three-stage model of human-AI attachment (Frontiers in Psychology, 2026).
