Does Conversation History Make Models Forget What They Know?

Language models often know a fact when asked about it directly, yet fail to use that same fact as one step inside a larger problem. Llama 3-70B states city time-zone offsets correctly about 90% of the time, but its accuracy falls to 25.4% when those offsets are intermediate steps in a combined time-and-place calculation. Putting the offset into the prompt restores performance (Around the World in 24 Hours). Separately, work on multi-turn dialogue finds that performance follows an inverted U as conversation history grows: some history resolves ambiguity, and more history buries the current turn (CORAL). These two findings have been studied apart. This proposal asks whether they are connected: does accumulated conversation history widen the gap between knowing a fact and applying it, so that a model hallucinates intermediate facts it would recall correctly in a clean prompt?

The core claim is testable and specific. We expect irrelevant history to suppress recall of facts stored in the model's weights ("parametric" facts) during multi-step reasoning. We expect this suppression to come in two separable forms. The first is dilution, which scales with history length. The second is interference, which scales with how semantically close the history is to the needed fact. We expect summarizing or pruning the history to fix the first form but not the second. If this holds, the right intervention is not better context compression. It is to run each recall step in a near-empty "clean-room" prompt and splice the recalled fact back into the reasoning as an explicit premise. That procedure turns a parametric recall problem into the in-context deduction problem that models already solve well (JEMHopQA).

What prior work establishes and leaves open

Three papers motivate the proposal by showing that the gap is a recall failure rather than missing knowledge. Empty Shelves or Lost Keys? profiles 2,150 Wikipedia facts. It finds that frontier models store 95–98% of them, that over 70% of GPT-5.2's factual errors are failures to recall stored facts, and that letting the model "think" before answering recovers 40–65% of these. That study is strictly single-hop and has no conversation history. The time-and-place study above shows that the failure appears when facts become intermediate steps. JEMHopQA shows that deduction is near-perfect once premises are supplied. None of the three varies the context around the question, so none can tell us whether history causes recall failure.

Three lines of work manage context, and all of them target facts that are present in the context rather than facts held in the weights. CORAL compares history strategies in conversational retrieval and finds that LLM summarization gives the best citation precision (31.1%). ACR trains a controller that rewrites dialogue history at turn boundaries and raises HotpotQA exact match from 21.7% to 35.1% while using 83.4% fewer tokens. COMPASS steers attention heads back toward the provided context when a hallucination detector fires. COMPASS pushes in the opposite direction from what we hypothesize: if history is the problem, more attention to context hurts parametric recall. We therefore treat it as a contrast case rather than a competitor.

Two papers supply our method and one constrains it. GRiD supplies the decomposition. It structures reasoning as a graph that separates "knowledge extraction" nodes, which recall facts from the weights, from "reasoning" nodes, which deduce from those facts. GRiD runs every node inside one shared single-turn prompt, so it never tests whether the surrounding context degrades the extraction nodes. CK-PLUG blends output distributions computed with and without context at every token. That is the closest existing mechanism to our clean room, and it is a real competitor. Its two contrasts with our method are that it operates per token at double decoding cost and that it was built for retrieval conflicts rather than dialogue history. Information Abundance Paradox constrains the design. It shows that long-context training shifts models toward reading context and away from their stored knowledge. That leads to a sharp secondary prediction: long-context variants of a model should lose more parametric recall under history than their short-context siblings.

Research Program

The first component is a controlled benchmark, which we call History-Loaded Composition. We start from the facts in WikiProfile and the GEOTEMP city–offset pairs. For each model, we keep only facts it recalls correctly in isolation. This filter makes every later failure a known-and-incorrect error by construction. We compose these facts into two- and three-step questions: GEOTEMP's time-and-place templates, plus chained WikiProfile relations of the form "the birthplace of the author of X." We then prepend conversation histories drawn from CORAL. CORAL generates conversations by walking Wikipedia title trees, which gives a natural relevance dial. A single-tree walk rooted near the target entity yields semantically close history. A dual-tree walk from an unrelated tree yields distant history.

The experimental design crosses three factors. History length ranges from zero to about 16 turns. History proximity is either near or far. Composition depth is one, two, or three steps. The primary metric is the conditional application rate: the probability of a correct composed answer given that every component fact was recalled correctly in isolation. A second metric is the intermediate hallucination rate: the fraction of cases where the model's stated intermediate fact differs from its own isolated answer. Both are hallucination measures, but they are defined relative to the model's own demonstrated knowledge, so they remove "it never knew" as an explanation.

The second component is a mechanism probe on open-weight models (Llama-3.1-8B, Qwen-2.5-7B and 14B, and Llama 3-70B). It tests whether attention mass flowing from the query-entity tokens into history predicts recall failure on an item-by-item basis. This is a correlational check in the spirit of COMPASS's context-reliance score, applied in reverse. We also compare the Phi-3 short- and long-context variants on identical items to test the context-addiction prediction.

The third component is the clean-room intervention. The model first writes a GRiD-style plan with explicit knowledge nodes. Each knowledge node is then answered in a separate prompt that contains only the node's sub-question, with references resolved by the model in the main context. The answer is inserted into the main context as a stated premise, and the reasoning nodes run as usual. To control cost, we evaluate a trigger that sends a node to the clean room only when the model's answer distribution for that node, computed in the full context, has high entropy. This applies CK-PLUG's entropy idea once per node rather than once per token.

Baselines

The naive reference point is the model answering with full raw history and no reasoning scaffold. History-length zero gives the clean-context ceiling for the same items.

The strongest compression baseline is LLM summarization of history, following CORAL, where it gave the best citation precision at 31.1%. We also include last-turn-only truncation, which CORAL compares under the name "Last Response." Summarization is the most direct test of whether dilution alone explains the gap. Every method uses the same base model, and the summarizer is that same model with a fixed prompt, so it costs no more than the intervention.

ACR is the strongest learned context manager. The file reports 35.1% exact match on HotpotQA against a 21.7% baseline. If released checkpoints are unavailable, we reproduce its Noise Filtering and State Abstraction operators by prompting and label the result as an approximation.

CK-PLUG, with dialogue history treated as "the context," is the strongest token-level competitor. It approximates a clean room through distribution blending. On KILT it moves memory recall on LLaMA3-8B across a range from 9.9% to 71.9%, against a fixed 42.1% for standard retrieval. No result on dialogue history or on our composed tasks appears on Lacuna, so we tune its blending weight on a held-out split with the same tuning budget we give our trigger threshold.

Reasoning-before-answering, in the form of plain chain-of-thought or extended thinking, is the baseline suggested by Empty Shelves' 40–65% recovery result.

The key ablation is GRiD-style knowledge nodes answered inside the full history. It shares every component of our method except the clean room, so any difference isolates the effect of removing context during recall.

The oracle upper bound injects the gold intermediate facts, mirroring GEOTEMP's injection experiment.

Beating the strongest baseline means two things. First, in the near-history condition, the method must achieve a higher conditional application rate than both summarization and CK-PLUG at matched or lower forward-pass cost, with confidence intervals from a paired bootstrap over items. Second, it must close at least half of the distance between full-history performance and the oracle.

Why not just…?

Why not just summarize or prune the history? This is the main threat, and the benchmark is built to settle it. If summarization recovers the gap in both near and far conditions, then the effect is pure dilution, the clean room adds nothing, and we will report that. Our prediction rests on the interference account: a summary of a near-topic conversation keeps exactly the neighboring entities that compete with the target fact. Summarization should therefore help far history much more than near history. CORAL's evidence concerns citation of facts that are in the context, so it does not speak to this question.

Why not just let the model think step by step? Empty Shelves shows that thinking recovers many stored facts, but only in clean single-hop prompts. The time-and-place study reports that chain-of-thought injected errors on simpler tasks and still collapsed to 25.4% on the hardest template. Reasoning traces are themselves context, and a wrong intermediate fact written early becomes conditioned truth for later steps. The thinking baseline, crossed with history length, will show whether thinking's recovery survives history or shrinks with it.

Why not just blend with a no-history distribution, as CK-PLUG does? This is the second real threat, because it targets the same quantity. Our method differs in granularity. Blending at every token mixes two distributions over the whole answer and doubles decoding cost. The clean room isolates only the recall steps and returns a discrete premise that later reasoning can cite. If CK-PLUG matches the clean room at equal tuning budget, the contribution reduces to the diagnosis plus a cheaper node-level trigger. A comparison of compute against accuracy will make that outcome visible.

Why not just retrieve the facts from an external source? Retrieval is a different intervention. It addresses facts the model never stored, whereas our items are filtered so that the model provably knows every component. Retrieval also adds context, which CK-PLUG shows can conflict with stored knowledge. We keep it out of scope.

Decisive Checks

The first check is whether the conditional application rate falls with history length at all. If the rate is flat out to 16 turns, the link between context windowing and recall failure is unsupported and the intervention is moot.

The second check is the dissociation between near and far history. If summarization recovers the far-history loss but not the near-history loss, while the clean room recovers both, the claim stands. If summarization recovers both, the interference account fails.

The third check is whether the clean room approaches the fact-injection oracle in the near-history condition, at lower cost than CK-PLUG.

The fourth check, a secondary one, is whether long-context Phi-3 variants show steeper losses than short-context variants. If the reverse holds, that would contradict the extension of context addiction from training to inference, and we would report it as such.