What is context rot in LLMs?

Context rot is the degradation in LLM output quality that occurs as input context grows longer. Chroma tested 18 frontier models and found every single one gets worse as more tokens are added, even before hitting the advertised context window limit.

At what token count do LLMs start to degrade?

Research shows significant degradation starts around 50K tokens for most models. A model advertising a 200K context window typically becomes unreliable between 50K and 130K tokens due to the lost-in-the-middle problem.

Which LLMs are affected by context rot?

All 18 frontier models tested by Chroma showed context rot, including GPT-4.1, Claude 4, Gemini 2.5, and Qwen3. No model was immune. The degradation is more severe on complex tasks than simple retrieval.

How do you fix context rot?

Current mitigations include RAG to avoid large contexts, chunked summarization, periodic context reset, and tools like Chroma CRoM. No complete solution exists — managing context length is the most effective current approach.

Does a larger context window mean better performance?

No. Context window size is a ceiling, not a performance guarantee. Adding more context often produces worse output, not better. Effective context management matters more than raw window size.

← Back to dashboard

clawsmith.com/signal/context-rot-chroma-llm-performance-degradation

📈 TrendsUnderservedLive

Context Rot: Chroma Tests 18 Frontier LLMs, Every Single One Degrades Past 50K Tokens

Chroma research shows all 18 frontier models degrade as context grows. Models advertising 200K context windows become unreliable at 50K to 130K tokens due to the lost-in-the-middle problem.

Product Idea from this Signal

A memory system that gives OpenClaw agents persistent recall that survives session crashes and context window degradation

49.8k ▲

OpenClaw agents lose their entire conversation history between sessions even when the files exist on disk. Silent daily session resets wipe agent memory without warning. Meanwhile, every frontier LLM degrades past 50K tokens (proven by Chroma across 18 models), meaning even within a session, agents progressively forget earlier context. ByteDance's OpenViking (19K stars in two weeks) proves massive demand for agent memory infrastructure. This tool gives OpenClaw agents a persistent, queryable memory layer that survives crashes, session boundaries, and context window limits by storing structured knowledge externally and injecting only relevant memories per turn.

DEVTOOLCLIAI-AGENTOPEN-SOURCE

CompetitiveView Opportunity →