Every way an AI keeps what it was told, what each one is good at, and where each one stops. Written to be fair to all of them. Updated as the field moves.
Memory here means three things at once: keep what was said, keep it exactly, and find it later — however much came after. Every model today gets one or two of the three. For each way of doing it: what it does, where it stops, and why it fails.
What it doesThe model reads through a context window — a buffer of recent text it can see all at once. Everything in the buffer is available. Hand it a document and it can reason over the whole thing. Sparse attention now makes the buffer much bigger and cheaper; the largest advertised windows reach twelve million tokens.
Where it stopsAt the edge of the buffer. When the session ends, or the buffer fills and old text is compressed or dropped, what was in it is gone. The model has no way to keep any of it.
Why it failsA window is a place to look, not a place to keep. Measured windows are also smaller than advertised — models lose what sits in the middle of a long context — and the cost of a conversation grows with everything ever said in it. A bigger or cheaper window moves the wall. It does not remove it.
What it doesKeeps the memory outside the model: a database, a document store, a memory app, a learned cache. A retriever pulls fragments into the window when the model needs them. It scales to any size of corpus and handles data that changes daily.
Where it stopsAt the seam. The model only sees what the retriever chooses. If the retriever misses, the model never knew.
Why it failsThe model never learned to use this memory, so retrieval quality is the ceiling on everything. Chunking breaks relationships — a term defined on page two, qualified on page twelve — before the model reads them. Most memory apps store a summary of what you said, not what you said; the number and the clause are gone. And it is a second system to run and keep in sync, while the model itself still forgets.
What it doesGives the model a state that updates as it reads: recurrent networks, state-space models, and newer designs that learn at test time or cache checkpoints of the state so it grows. Linear cost, cheap on long streams, good at anything a summary can serve.
Where it stopsAt exactness. It can tell you roughly what was there. It cannot give the words back.
Why it failsA fixed-size state absorbing an unbounded sequence is a compression, and compression is lossy. Exact recall degrades as the sequence grows; transformers beat these models at copying; even the growing-memory versions report that attention still wins at recall. The hybrids in production keep attention layers for one stated reason: to preserve the retrieval the recurrent layers cannot.
What it doesPuts a lookup table inside the model — tens of billions of parameters of it — so facts seen in training can be pulled up in constant time. Models with a memory table do better per unit of compute on knowledge and reasoning.
Where it stopsAt the training set. Nothing said at inference time is written to it.
Why it failsIt makes a model know more. It does not make a model remember.
What it doesTakes a trained model and adds a memory to it. For some parts this works: swapping dense attention for sparse attention trades one computation for a cheaper one the model already knows how to use.
Where it stopsAt the memory itself. A model trained without one has no use for one; nothing in its training ever asked it to write or read.
Why it failsThe memory fights the model, and harder the bigger the model is. In one 2026 study, an exact memory grafted into a one-billion-parameter model cost two percent of its quality; at four billion, eleven percent; at twelve billion, forty-four percent. Our own graft onto a trained one-billion-parameter model ended in a clean negative result. Memory has to be trained in from the first step, together with the language.
Each of these keeps part of what it was told, for a while, in some form. None keeps what was said, exactly, past the window, inside the model. Current models do not remember — not truly.
Kymnera is building one that does.
v1 — 7 September 2026. First published. Corrections welcome at the address above.