Context window hygiene

Context Window Hygiene

A bigger context window is not a memory upgrade. It is a more expensive way to be wrong.

The model attends to the prompt it actually sees. Long contexts hide signal under noise — middle-of-context recall degrades, key facts get out-ranked by chatter, and cost scales linearly with every replayed turn. Hygiene is about removing what doesn't earn its place, not adding more.

When to use this skill

• The user is running a long agentic loop (tool-use loop, multi-turn task).
• The user reports the agent "forgetting" mid-run, or contradicting earlier turns.
• The user is paying for context tokens that grow linearly with conversation length.
• The user is considering "just use a 1M-token window" as a solution to anything.

What enters context — name the categories

Every token in the prompt is one of:

• System prompt — fixed instructions. Should be cache-aligned.
• Skill / tool definitions — fixed shapes. Cache-aligned.
• Conversation history — previous turns. Pruning candidates.
• Tool results — output of tool calls. Often the biggest growth vector.
• Retrieved context — RAG hits, memory recall. Bounded budget.
• Current turn — user input.

Distinguish them. They have different lifetimes, different cache behaviour, different prune rules.

Decision flow

1. Is context growth driven by tool results? → yes (it usually is) → summarise large tool results before they enter history.
2. Is context growth driven by long conversation? → yes → compaction at turn N, not at the context limit.
3. Are old turns still being read? → no → drop them. The model can't read what isn't there, but it pays attention to what is.
4. Is retrieved context unbounded? → yes → set a recall budget (≤K items, ≤T tokens) — see [[memory-design]] and [[rag-vs-context-engineering]].
5. Is the system prompt mutating per call? → cache is dead. Stabilise the prefix.

The hygiene rules

1. Cache-aligned order. Stable prefix first (system → tools → fixed examples), variable suffix last (history → current turn). Cache miss is paid per request; alignment turns long history from a cost into a free re-read.
2. Summarise tool results that exceed a threshold. Raw 50KB JSON in history poisons every subsequent turn. Summarise with a typed schema before storing in conversation.
3. Compact at fixed boundaries, not at the cliff. Compact at turn 10, 20, 30 — not when you hit the context limit. Predictable cost, predictable behaviour.
4. Lossy compaction is fine; lying compaction is not. Drop tool-result chatter; keep facts the agent decided on. Never invent context that wasn't there.
5. Recency + saliency, not recency alone. Drop the most-recent-but-irrelevant before the older-but-load-bearing. The model uses what's there, not what was there.

Anti-patterns to flag immediately

• "Just use a 1M-token window." Solves cost-per-call by raising every-call's cost. Doesn't solve middle-of-context recall degradation. Doesn't solve attention drift.
• Full-history replay with no pruning. Cost grows quadratic-ish in real workloads (every call pays for every prior turn).
• Tool results in history verbatim. Raw HTML, full DB dumps, base64 — all attend, all confuse, all cost.
• Mutating system prompt per call ("Current time is X"). Kills cache; cost balloons; cache hit rate hides in the metrics.
• Compacting at the cliff. Hitting 95% context, panicking, summarising — produces lossy compaction at the worst moment.
• Free-text summary memory. Summaries written as paragraphs lose precision; key facts blur.
• No measurement of attention. "It feels like the agent forgot" with no diagnostic from the trace.
• Confusing context window with memory. See [[memory-design]] — they aren't the same.

Diagnostic checklist when the agent "forgets"

1. Was the fact ever in the context window of the failing call? (Read the trace.)
2. Was it in the first quarter or the middle of the prompt? Middle-of-context recall degrades; surface it.
3. Was it surrounded by noisy tool output that out-competes for attention?
4. Did a compaction step drop or paraphrase it?
5. Is the fact in a documented format the model is following, or buried in prose?

Questions to ask the user

1. What's in your prompt prefix vs suffix — is it cache-aligned?
2. How big is the average tool result? Are they summarised before re-entry?
3. What's the compaction policy — turn boundary, token threshold, or "when we hit the limit"?
4. What's the recall budget for memory / RAG hits in a single turn?
5. When the agent "forgets", can you trace what was in context at that call?
6. Is your system prompt stable across calls, or does it change per-request?

The hard line

Every token in context must be load-bearing on this turn's decision. If a token isn't, it's noise — and noise costs you in money, latency, and quality.

Why this exists

The "1M-token context" pitch made everyone forget that attention degrades with length, recall degrades in the middle of long contexts, and cost scales linearly with every token shipped. Hygiene — pruning, summarisation, cache alignment, compaction at boundaries — buys back the coherence that long contexts lose, at a fraction of the cost.

References

• references/prefix-order.md — cache-aligned prompt order, system → tools → examples → history → turn.
• references/compaction.md — when, how, and what to drop; lossy vs lying; summarisation templates.
• references/tool-result-summarisation.md — schemas for compacting big tool outputs before they enter history.

Related skills

• Hygiene is not memory — [[memory-design]].
• Hygiene is not retrieval — [[rag-vs-context-engineering]].
• Hygiene drives cache hit rate — [[prompt-caching]].
• Hygiene drives cost — [[agent-cost-modeling]].
• Hygiene drives latency — [[latency-budgeting]].
• Long runs need observability — [[agent-observability]].
• Tool result size is set in [[tool-use-schema-design]].