By turn 50 or 60 of a Claude conversation you notice it starting to forget: naming conventions you set 10 turns ago, options you already rejected creeping back, sometimes contradictions within a single answer. This isn’t laziness — it’s how the context window actually works. Long chats become unstable; the trick is knowing when to reset.
Claude’s window (Sonnet 200K, Opus 200K-1M) is large in name, but attention to “middle content” drops sharply compared to the start and end (the lost-in-the-middle effect). The longer the chat, the more your early decisions get neglected.
Common causes
Ordered by hit rate, highest first.
1. Approaching window cap — old messages silently truncated
When input crosses ~200K tokens, the UI auto-trims from the middle without telling you. What gets cut is the “not-recent, not-ancient” band — exactly where your early design decisions live.
How to spot it: Settings → Usage — if input tokens are at 180K-200K, content is almost certainly being trimmed.
2. Lost-in-the-middle: model favors head and tail
Even when the window isn’t full, recall drops sharply for the middle 30-70% of the conversation. By turn 40, decisions from turns 15-25 are in the forgetting band.
How to spot it: Ask Claude to recite a specific decision from turn N. Fuzzy or wrong = this is it.
3. You gave contradictory instructions over time
Turn 5: “use snake_case.” Turn 20, offhand: “camelCase is fine too.” Turn 40: Claude is mixing both — because both “boss-said” quotes are in context.
How to spot it: Re-read your own prompts. Any self-contradictions?
4. A large Artifact / file is hogging attention budget
A 5,000-line artifact in the conversation is re-read every turn, leaving less attention for actual reasoning.
How to spot it: Is there an unclosed large Artifact in the chat?
5. Multiple topics interleaved
One chat discussing the design of 3 different features, each with its own decisions. Claude blends details across them.
How to spot it: Is the chat single-topic, or A + B + C mixed?
6. Extended Thinking accumulation
Thinking-mode reasoning is hidden but still occupies context. Many turns of thinking dilute the window.
How to spot it: Thinking mode on + long conversation.
Shortest path to fix
Step 1: Have Claude list “all decisions so far”
In the messy chat:
List every decision we've made so far:
- naming conventions
- architectural choices
- approaches we rejected
- open questions still undecided
Markdown list. No new commentary.Copy that list. Fix anything wrong by hand — this is your decision summary.
Step 2: Open a new chat, lead with the summary
First message in new chat:
We were discussing project X. Key decisions:
[paste the list]
Continue from here. Do not re-debate settled items.Fresh attention beats fighting a polluted long chat.
Step 3: Use file attachments for big material
For code or docs over ~200 lines, upload as an attachment instead of pasting in triple backticks. Attachments are referenced on demand internally, friendlier to attention.
Step 4: One topic per chat
Discussing 3 features? Three chats. Treat each chat like a PR or issue, not a stream-of-consciousness log.
Step 5: Periodically /compact-style reset
Claude Code: /compact compresses the session. Web has no equivalent command, but you can manually:
In 500 words, summarize what we have so far:
1. Task background
2. What's done
3. What we're working on now
4. What's next
I'll start a fresh chat with this pasted in.Step 6: Don’t change foundational rules mid-chat
If you actually want to switch snake_case → camelCase, don’t drop it casually in turn 40 — start a new chat and re-establish the rule. Changing in-flight leaves “ghosts” of the old rule in attention.
Prevention
- Set a hard 20-30 turn threshold — proactively summarize and open a new chat, don’t wait until it’s already broken
- For long projects, maintain a “decisions list” in Project Knowledge or external doc, paste at chat start
- One topic per chat — never discuss features A, B, C in a single thread
- Use attachments / Projects Knowledge for large material, not direct paste
- Pin important decisions (architecture, conventions) in the SYSTEM PROMPT, not in chat
- For long Claude Code sessions, proactively
/compact
Related
- Claude usage limit
- Claude inaccurate answers
- Claude beginner guide
- Claude prompt best practices
- Claude Projects
Tags: #Claude #Debug #Troubleshooting