Why longer conversations get worse

Three hours into a build, one developer filed it as plainly as a bug can be filed: “During a long session (~3+ hours), Claude Code’s quality of responses and decision-making degrades significantly.” It doesn’t take three hours to notice. On Hacker News, another developer traced the same drift to something more specific than fatigue: “The LLM context collapse happens well before you hit the token limits, and things like ground rules and whatnot stop influencing the LLMs outputs after a couple tens of thousands of tokens.”

Neither report mentions running out of context. Neither mentions /compact firing. The transcript is still completely there. The model has just gotten worse at using it.

The most direct measurement of this comes from Chroma Research’s “Context Rot” study (Hong, Troynikov & Huber, July 2025): 18 models — GPT-4.1, Claude 4, Gemini 2.5, Qwen3 among them — run across 194,480 total LLM calls. The degradation isn’t a straight line down as input grows; it comes in non-uniform cliffs. When the fact a model needs sits at positions 5 through 15 of a 20-item context, accuracy drops more than 30 points. And it isn’t only about distraction: with competing content masked out entirely, the accuracy floor still falls 7.9 percentage points from length alone.

A second study — Microsoft Research and Salesforce, “LLMs Get Lost in Multi-Turn Conversation” — measured the same decay from the outside, across 200,000-plus simulated conversations. As a task’s complexity rises, responses grow 20 to 300 percent longer, with proportionally more hallucinated detail padded in. And once a model commits to a wrong assumption early in a conversation, it doesn’t self-correct — it doubles down, defending the wrong turn instead of reversing it.

Put those together and a three-hour session is failing two ways at once: the detail it needs from an hour ago is statistically likely to be sitting in the worst position to retrieve, and the wrong guess it made twenty minutes in is now something it’s actively defending rather than reconsidering.

Push through it.This isn’t noise that averages out with more turns — it’s a measured cliff, and per the multi-turn study, a model that takes a wrong turn doesn’t wander back on its own. It defends the mistake. More turns on the same long context tends to compound the error, not dilute it.

Restart every 30–45 minutes.It genuinely resets the clock. The cost is that you re-explain everything by hand at the top of every session, which is you doing the model’s remembering — fine for a handful of facts, a real tax once a project runs for weeks.

A bigger context window.It buys room before the cliff, not a flatter curve. Chroma’s 7.9-point floor drop held even with distractor content masked out — length alone did that, independent of what else was competing for attention. A wider window is a wider middle. See vs. a bigger context window.

Keep repeating the instructions.The Hacker News report above is the ceiling here in one line: ground rules stop influencing output “after a couple tens of thousands of tokens” — including rules restated mid-session, which are now themselves aging content, subject to the same decay.

None of this moves the cliff. It changes how often a real session runs into it, and it costs nothing to adopt.

Keep sessions short and single-purpose, on purpose.One task, one session, then stop — before the cliff, not after it. This is deliberate hygiene, not a product feature, and it’s the cheapest lever available.

Put load-bearing facts at the start or the end, never buried mid-task. The measured strong positions are the ends of a context, not its middle — so the instruction the whole session depends on belongs in the system prompt or the most recent turn, not paragraph six of something pasted in an hour ago.

Trim what you feed it.A full log dump or an entire file pasted “just in case” is dead weight sitting in the exact stretch that’s hardest to retrieve from. Less irrelevant content means less of a middle for the answer to get lost in.

Write the durable stuff down somewhere external. A CLAUDE.md, a decisions log, a design doc — anything the next session reads instead of re-deriving. It’s what makes restarting cheap instead of costly.

These all buy distance from the cliff. A session that runs genuinely long enough reaches it regardless of how carefully it was packed — which is the ceiling every one of the above shares.

Both figures below are from the papers themselves — external, checkable, and not ours.

The companion figure — 20 to 300 percent longer responses as complexity rises, and no self-correction once a model commits to a wrong turn — is from Microsoft Research and Salesforce’s multi-turn study, measured across 200,000-plus conversations (arxiv.org/abs/2505.06120).

The cliff above is a property of how much a model has to reason across in one context, not a property of how many turns a session has run. A session that reads a compact, addressable brief each turn — instead of attending across an ever-larger transcript — isn’t fighting the same slope. One production session on Vinculum ran 11,232 turns, and the fresh material it reasoned over per turn stayed at roughly 116 tokensthe whole way through — turn 11,232 asked no more of it than turn 1 did. That’s a different shape than a transcript that grows with every exchange until something has to summarize it away or start eating cliffs.

It doesn’t repeal the research above — a large-enough context, cached or not, still has a middle. It changes how much of a session’s history has to sit in that context at all. This is model-agnostic: the same idea works whichever model is doing the reasoning.