The finding that a model's ability to recall a stored fact depends on how closely the query matches the context in which the fact was originally encoded during training.
Context-Dependent Recall is the observation that a fact stored in a model's parametric memory may or may not be retrievable depending on how the query is phrased and directed. The fact is encoded in the weights, but the retrieval pathway from the prompt to that fact is not equally strong from every angle.
Calderon and Yona (Google Research, 2026) measured this directly. They evaluated 13 LLMs on 2,150 Wikipedia-derived facts and found that frontier models encode 95-98% of facts in their weights but fail to recall 26-34% of them. The bottleneck is not storage but access: recall is tightly coupled to the conditions under which a fact was learned.
A model trained on "A leads to B" often cannot reverse the relationship. When asked "What is B?", it retrieves the answer. When asked "What leads to B?", the pathway from the category back to the entity is too weak. In multiple-choice format (recognition rather than generation), models answer the reverse question correctly, which confirms the knowledge is encoded. The failure is in open-ended recall, not in storage.
The encoding gap between popular and rare facts is modest (over 90% of long-tail facts are encoded). The recall gap is much larger, dropping 20-35 percentage points for niche information. Rare facts are present in the weights but harder to reach when the prompt diverges from training-time context.
When models are given intermediate reasoning steps before answering (chain-of-thought, "thinking" modes), they recover 40-65% of encoded-but-inaccessible facts. For facts not encoded at all, thinking recovers only 5-15%. This suggests thinking acts primarily as a recall-facilitation mechanism, constructing associative bridges across the weights to reach facts that direct generation cannot.
Context-dependent recall reframes a class of factual errors. A hallucination where the model confidently states an incorrect fact about a brand may not be a gap in training data. The correct fact may be fully encoded but unreachable from the direction the user approached it. The model fills the gap with statistically probable tokens because generation does not distinguish between "not stored" and "stored but not reachable from here".
