Skip to content

Key Challenges in Context Management

Overview

As AI agents handle longer tasks and accumulate tool call feedback, the context window becomes a critical bottleneck. Context quality degrades in predictable ways — understanding these failure modes is the first step to engineering around them.

The term "context engineering" was popularized in part by Cognition (builders of Devin), who described it as "effectively the #1 job of engineers building AI agents." Anthropic echoed this: "Agents often engage in conversations spanning hundreds of turns, requiring careful context management strategies."

The Five Context Failure Modes

These failure modes were outlined by Drew Breunig in How Long Contexts Fail and referenced widely across the community:

Context Rot

Degradation of model performance as context length increases. Studies using needle-in-a-haystack benchmarks show that as token count grows, a model's ability to accurately recall information from that context decreases — even within the stated context window limit.

Anthropic's engineering team describes the underlying mechanism: the transformer architecture creates n² pairwise relationships for n tokens. As context grows, the model's ability to capture these relationships gets stretched thin. Models also have less training experience with very long sequences, meaning performance degrades gradually rather than hitting a hard cliff.

The Pokémon-playing Gemini agent team observed this empirically: beyond ~100,000 tokens, the agent began favoring repetition of past actions over synthesizing novel plans — a clear sign of context distraction caused by context rot.

Context Poisoning

When a hallucination or other error enters the context and is subsequently treated as ground truth. Because agents reference prior context to make decisions, a single bad output can propagate and compound across many subsequent steps.

Manus's engineering team notes this as a core reason to leave errors visible in context rather than cleaning them up: "Erasing failure removes evidence. And without evidence, the model can't adapt." Hiding errors prevents the model from updating its beliefs away from the failed approach.

Context Distraction

When the context grows so large that the model over-focuses on accumulated history, neglecting what it learned during training. This manifests as agents that repeat patterns from their history rather than reasoning freshly about the current situation.

This is distinct from context rot (general degradation) — distraction specifically refers to the model being pulled toward in-context patterns at the expense of trained knowledge and reasoning.

Context Confusion

When superfluous or irrelevant information in the context influences the model toward lower-quality responses. Tool descriptions are a common source: research shows that above ~30 tools, descriptions begin to overlap and cause selection errors. Above 100 tools, failure rates become very high.

The "Less is More" paper demonstrated that Llama 3.1 8b fails a benchmark when given 46 tools but succeeds with only 19 — the issue is context confusion, not context window limits.

Context Clash

When parts of the context contain conflicting information or when new tools and data contradict earlier context. This is especially problematic in multi-agent systems where subagents operate on different assumptions.

Cognition's Walden Yan describes this as a core failure mode of naive multi-agent architectures: "Actions carry implicit decisions, and conflicting decisions carry bad results." When subagents work in parallel without shared context, their outputs reflect incompatible assumptions that the orchestrator must then reconcile.

Impact on Agent Performance

Failure Mode Primary Symptom Common Trigger
Context Rot Reduced recall accuracy, long-range reasoning errors Very long conversations, many tool calls
Context Poisoning Repeated mistakes, compounding errors Hallucinations left uncorrected in history
Context Distraction Pattern repetition, reduced creativity Context > ~100K tokens
Context Confusion Wrong tool selection, irrelevant responses Too many tools, noisy message history
Context Clash Inconsistent outputs, contradictory decisions Multi-agent parallelism without shared context

Memory vs. Context: The Fundamental Distinction

Most teams treat memory and context as the same thing. They are not, and the distinction determines whether a system scales.

Aspect Context (Working Memory) Memory (Long-term Storage)
Cost Expensive: every token costs money, 10× more if uncached Cheap: storage costs are negligible
Size Limited: by the context window Unlimited: store millions of documents
Speed Immediate: direct influence on response Slow: requires retrieval before use
Persistence Volatile: cleared between sessions Persistent: survives across sessions
Influence Direct: affects every model decision Indirect: must load into context first
Performance Degrades after ~30K tokens No degradation: just retrieval overhead

The 100:1 rule: For every token an agent generates, it processes roughly 100 tokens of input. Context management is therefore the dominant cost factor in production agent systems.

What to keep in context vs. store in memory:

Keep in Context Store in Memory
Current task objectives and constraints Historical conversations and decisions
Recent tool outputs (last 3–5 calls) Learned patterns and preferences
Active error states and warnings Large reference documents
Immediate conversation history Intermediate computational results
Currently relevant facts Completed task summaries

Four Types of Context

Agents handle four distinct kinds of context, each with different failure characteristics:

  1. Instructions context — system prompts, few-shot examples, behavioral guidelines. Tends to be stable but can grow large. Poor instruction context leads to inconsistent behavior.
  2. Knowledge context — facts, memories, retrieved documents, user preferences. Dynamic; changes based on task. Can quickly overwhelm the context window.
  3. Tools context — tool descriptions, feedback from tool calls, error messages. Every model degrades with tool count (Berkeley Function-Calling Leaderboard). Tool context management is critical.
  4. History context — past conversations, previous decisions, learned patterns. Valuable for continuity but can introduce contradictions and outdated information.

Empirical Evidence for Each Failure Mode

Context Poisoning — DeepMind Pokémon Agent

The DeepMind team documented this with their Pokémon-playing Gemini agent. The agent misidentified the game state once, and that error got embedded in the goals section. Because goals are referenced at every decision point, the false information reinforced itself. The agent spent dozens of turns pursuing impossible objectives and was unable to recover because the poisoned context kept validating the error.

Domain cascade pattern: - Customer service: Misidentifies product model → wrong troubleshooting → references incorrect manual → suggests incompatible accessories - Code generation: Imports wrong library version → deprecated APIs → incompatible syntax → cascading type errors - Research: Misunderstands scope → searches wrong domain → cites irrelevant sources → draws incorrect conclusions

Recovery is nearly impossible once poisoning sets in. Clearing the entire context works but loses all progress. The agent trusts its own context more than external corrections.

Context Distraction — Gemini 2.5 Team

The Gemini 2.5 team observed this when their agent's context grew beyond 100,000 tokens. The agent stopped generating novel solutions and started repeating actions from its vast history, even when those actions didn't fit the current problem. Databricks researchers found that when models hit their distraction threshold, they often ignore instructions entirely and just summarize whatever's in the context.

Context Confusion — Berkeley Function-Calling Leaderboard

A quantized Llama 3.1 8B with a 16,000-token context window was given 46 tools from the GeoEngine benchmark (~3,000 tokens of context). The agent had plenty of room but failed completely. When tools were reduced to 19, it succeeded. The problem is signal-to-noise ratio: each additional tool makes existing tool descriptions less distinct.

Context Clash — Microsoft and Salesforce Research

Researchers took standard benchmark tasks and spread them across multiple conversation turns instead of single prompts. Performance dropped 39% on average across all tested models. When models take a wrong turn in a conversation, they get lost and do not recover — the context contains both the error and the correction, and the agent can't consistently prioritize the correction over the error.

Context clash vs. context poisoning: Poisoning involves a single error that propagates. Clash involves multiple contradictory pieces of information competing for influence.

Context Size Recommended Approach
< 10K tokens Simple append-only approach with basic caching
10K–50K tokens Add compression at boundaries with KV-cache optimization
50K–100K tokens Implement offloading to external memory with smart retrieval
> 100K tokens Consider multi-agent isolation architecture

Common Anti-Patterns

Anti-Pattern Consequence Correct Approach
Ignoring error messages Agents repeat the same mistakes Keep error context to enable learning
Modifying previous context Breaks cache validity and multiplies costs Use append-only designs to preserve cache
Over-engineering early Adds complexity that models will outgrow Start simple and add complexity when needed
Loading all tools upfront Degrades performance as tool count increases Use dynamic tool selection based on task
Aggressive pruning without recovery Loses information permanently Combine compression with offloading

See Also

References