Researchers who profiled 13 large language models across more than 4 million responses report that the frontier systems have very nearly stopped forgetting: GPT-5 and Gemini 3 encode 95–98% of the facts tested against them. What those models cannot reliably do is get the facts back out. Asked directly, without reasoning, they fail to produce 26–34% of what already sits in their parameters. Extra computation at inference time recovers 40–65% of those misses. The failure the industry keeps treating as missing knowledge is, in the best models, mostly a retrieval problem inside the model itself.
The study's central move is to split one thing the field usually measures as a single number. Encoding means the fact is written into the weights, tested by asking the model to continue the original passage it most likely trained on. Recall means the model answers questions about that fact reliably, across different phrasings and from different directions. Ordinary benchmarks score the two failures identically. The fixes are opposite. A fact that was never encoded is a pretraining problem, solved by more data or more capacity. A fact that is encoded but unreachable is a post-training and inference problem, and no amount of pretraining spend touches it.
The authors work the distinction through one example: the band Oasis played its first show at a club called the Boardwalk. Five profiles follow. Direct recall — the fact is stored and comes out immediately. Encoding failure, which they call empty shelves — the model can neither continue an encyclopedia sentence about the band's early gigs nor answer questions about them. Recall failure, or lost keys — the model happily continues the original training text about Oasis but will not say where the band first played, even given time to think. Recall through reasoning — the fact is stored and unreachable by direct generation, but a chain of thought about the band's early years in Manchester surfaces it. And inference without encoding — the specific fact was never stored, but the model guesses correctly by combining what it knows: Oasis formed in Manchester, the Boardwalk was a well-known 1990s music venue, new bands played there.
The measurements come from WikiProfile, a benchmark of 2,150 facts drawn from Wikipedia, each probed in formats ranging from exact continuation of the source context to multiple choice. The researchers liken the recall gap to having a word on the tip of your tongue: the information is there, and it arrives only after extra effort and a mental reconstruction of context.
The most useful result is what happens when you scale. On Gemma 3, going from 1 billion to 27 billion parameters cut encoding failures from 85% to 23% — exactly what the scaling story predicts. Recall failures went the other way, reaching 40% without reasoning. Scale fixes storage and does nothing for access, so the errors migrate rather than disappear. The more a model memorises, the larger the pile of knowledge that is encoded but unreachable.
Phrasing decides a great deal. Rare facts are encoded at roughly the same rate as popular ones, but the recall gap between obscure and widely known facts exceeds 25% in frontier models. Reverse questions are similarly brittle: a model that states without hesitation that Oasis debuted at the Boardwalk may be unable to name which band debuted at that club — and will then pick the right answer when the same question is offered as multiple choice. Recognition survives where generation fails, which is a strong sign the knowledge is present and the access path is not.
The practical consequence is the part worth arguing about. Nitay Calderon, a research scientist at Google, makes the point that teams reflexively reach for larger models or more data when a fact the model already holds is not the problem, and that many of the jobs companies hand to retrieval-augmented generation the model can already do from memory. That is a real and uncomfortable claim: if a fact is genuinely absent, retrieval earns its latency and its per-query cost; if it is encoded but not recalled, a vector database is an expensive way to avoid diagnosing the failure. My read is that this lands hardest on the enterprise architectures built in the last two years, where "add RAG" became the default response to any hallucination without anyone first checking which of the two failures they were looking at.
What the paper cannot yet give those teams is the thing they would actually need. Reasoning recovers 40–65% of unreachable facts but is required for only 10–20% of them, so switching it on for every query burns compute to no purpose. Routing queries to deeper reasoning on demand is the obvious answer, and the researchers concede current models are not reliable at noticing that an answer is about to be wrong — the model would have to recognise its own uncertainty before it generates. Google is working on approaches described as confident uncertainty for precisely this. Until something like that ships, the actionable advice reduces to the cheap end of the list: rephrase the prompt and retry, add intermediate context, ask for reasoning first and the answer second, and build generate-then-verify pipelines that exploit the recognition-beats-generation asymmetry by having the model check its own claims as a separate step.
Notably absent from the framing is how far any of this travels. Encoding is measured by continuation of source text the model probably trained on, and the corpus is Wikipedia — about the most heavily represented text in any pretraining mix. Calderon says as much in the caveat: the pipeline is built to run on a new corpus, but where Wikipedia's problem was recall, a specific domain may simply not contain the facts in the model at all. That inverts the headline finding exactly where the money is. Teams can check for themselves — WikiProfile is on Hugging Face with the exact prompts used to build it, and a full profile of a frontier model costs about $500, less if you drop the multiple-choice variants or sample each question fewer times. Five hundred dollars to find out whether your retrieval stack is solving your problem is the cheapest experiment in this entire story.
The finding that should unsettle people is not that models forget. It is that every increment of scale grows the store of knowledge that is present and unreachable, and the one tool that would open it — a model that knows when it needs to think harder — is the piece nobody has shipped.