By turn 50 or 60 of a long Claude chat you notice it forgetting: naming conventions you set 10 turns ago, options you already rejected creeping back, sometimes contradictions inside a single answer. This is not laziness. It is how the context window behaves as it fills, and the fix is almost always to reset rather than to keep pushing the same thread.
Fastest fix (do this first): in the messy chat, ask Claude to list every decision made so far, copy that list, open a new chat, and paste the list as your first message. Fresh attention on a clean summary beats fighting a polluted 60-turn thread every time. The rest of this page explains which specific cause you are hitting so the reset actually sticks.
One myth to kill up front: the window is huge now. As of June 2026, Claude Sonnet 4.6 and Opus 4.7 both have a 1M-token context window (made generally available in March 2026, up from the old 200K). Size is not the problem. The problem is that attention to the middle of a long conversation drops sharply versus the start and end — the “lost in the middle” effect — and fidelity to early instructions starts slipping once the window is roughly 60-70% full, well before any hard limit.
Which bucket are you in?
Work down this table; the top rows are the most common.
| Symptom you see | Likely cause | Fastest check |
|---|---|---|
| ”Hit the maximum length for this conversation” banner; replies feel like they ignore early setup | Window near full, oldest/middle messages summarized or trimmed | Look for the length-limit banner near the composer |
| Recalls recent and very first messages fine, but mangles a decision from the middle | Lost-in-the-middle | Ask it to recite a specific decision from turn N |
| Mixes two conventions you both mentioned | You gave contradictory instructions over time | Re-read your own prompts for a self-contradiction |
| Reasoning feels “thin,” lots of context spent | A large Artifact or pasted file is hogging the budget | Is there a 1,000+ line Artifact or paste in the thread? |
| Blends details across unrelated things | Multiple topics interleaved in one chat | Is the chat single-topic or A + B + C mixed? |
| Slow drift over many thinking-heavy turns | Extended Thinking accumulation | Thinking mode on plus a very long chat? |
Common causes, in detail
1. Window near full — old messages get summarized or trimmed
When a conversation approaches the 1M-token limit, the claude.ai web app no longer just hard-stops. On paid plans, if code execution is enabled, Claude now automatically summarizes earlier messages to keep going, and your full history is preserved so it can still be referenced. That is helpful, but a summary is lossy: precise early design decisions get compressed into a sentence, which is exactly where your early choices live.
You may still hit a hard wall on edge cases (for example, a very large first message), shown as: Claude hit the maximum length for this conversation.
How to spot it: watch for the length-limit banner near the message box, or a notice that earlier messages were summarized. In Claude Code, run /context to see exactly how full the window is.
2. Lost in the middle — the model favors head and tail
Even when the window is far from full, recall drops for the middle band of a long conversation. By turn 40, decisions from turns 15-25 sit squarely in the forgetting zone. Research and Anthropic’s own guidance both note that quality degrades as context fills, typically becoming noticeable past 60-70% capacity.
How to spot it: ask Claude to recite a specific decision from a known earlier turn. Fuzzy, partial, or wrong recall confirms it.
3. You gave contradictory instructions over time
Turn 5: “use snake_case.” Turn 20, offhand: “camelCase is fine too.” Turn 40: Claude mixes both, because both “you said” quotes are still in context with equal weight.
How to spot it: scroll back and re-read your own prompts. Any place you softened or reversed a rule is a candidate.
4. A large Artifact or pasted file is hogging the budget
A 5,000-line Artifact or a giant pasted file is re-read on every turn, spending attention that could go to reasoning.
How to spot it: look for an open, large Artifact or a huge code paste in the thread.
5. Multiple topics interleaved
One chat designing three 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 braided together?
6. Extended Thinking accumulation
Extended Thinking reasoning is hidden from you but still occupies context. Many thinking-heavy turns dilute the window faster than the visible text suggests.
How to spot it: Thinking/Extended Thinking on plus a long conversation.
Shortest path to fix
Step 1: Have Claude list every decision so far
In the messy chat, send:
List every decision we have made so far:
- naming conventions
- architectural choices
- approaches we rejected and why
- open questions still undecided
Output a Markdown list. No new commentary.Copy the list and fix anything wrong by hand. This is now your decision summary, and it is more reliable than whatever Claude can still recall from the middle of the thread.
Step 2: Open a new chat, lead with the summary
We were working on project X. Key decisions so far:
[paste the list]
Continue from here. Do not re-debate settled items.Fresh attention on a clean summary beats fighting a polluted thread. Anthropic’s own guidance is to start a new conversation roughly every 15-20 messages on heavy work and carry a summary forward.
Step 3: Use file attachments for big material
For code or docs over ~200 lines, upload them as an attachment (or put them in Project Knowledge) instead of pasting inside triple backticks. Attached material is referenced on demand rather than re-read in full every turn, which is friendlier to attention.
Step 4: One topic per chat
Three features means three chats. Treat each chat like a PR or an issue, not a stream-of-consciousness log. This single habit prevents most cross-topic blending.
Step 5: Reset on a schedule, not after it breaks
In Claude Code, you have real tools for this:
/clear— wipe context entirely between unrelated tasks (no token cost)./compact— summarize and shrink the session; add focus with/compact Focus on the API changes.Esc Escor/rewind— pick a checkpoint and choose Summarize from here or Summarize up to here to condense part of the thread while keeping the rest in full./context— check how full the window is so you act before degradation, not after.
On the web there is no /compact command, but you can reproduce it manually:
In 500 words, summarize where we are:
1. Task background
2. What is done
3. What we are working on now
4. What is next
I will paste this into a fresh chat to continue.Step 6: Do not change foundational rules mid-chat
If you genuinely want to switch snake_case to camelCase, do not drop it casually at turn 40. Start a new chat and re-establish the rule cleanly. Changing it in flight leaves “ghosts” of the old rule in context, and the model keeps half-honoring both.
How to confirm it is fixed
After resetting, send one targeted test before you continue real work:
Restate, in one line each, our naming convention, our architecture choice,
and one approach we rejected.If all three come back correct and specific, the new chat has clean context and you are good to go. If any are still fuzzy, your summary was incomplete — add the missing decision to it and re-paste.
Prevention
- Set a hard 20-30 turn threshold: summarize and open a new chat proactively, before it breaks.
- For long projects, keep a living “decisions list” in Project Knowledge or an external doc and paste it at the start of each chat.
- One topic per chat, always.
- Use attachments or Project Knowledge for large material, never a raw paste.
- Pin durable decisions (architecture, conventions) in a project’s custom instructions, not in chat prose where they can drift.
- In long Claude Code sessions, run
/contextto watch the fill and/compact(or/clearbetween tasks) before quality slips.
FAQ
Does the 1M-token window mean long chats no longer break? No. A bigger window delays the hard limit but does not fix lost-in-the-middle. Fidelity to early instructions still degrades as the window fills, typically becoming noticeable past 60-70% capacity, so resetting periodically still matters as of June 2026.
What does “Claude hit the maximum length for this conversation” mean? You have filled the conversation’s context window. Start a new chat to get a fresh window. Ask the old chat for a summary first, copy it, and paste it as the first message in the new one so you do not lose context.
Will I lose my history if Claude summarizes the conversation? No. On paid plans with code execution enabled, Claude summarizes earlier messages to keep going but preserves the full history so it can still reference it. The risk is precision, not data loss — fine-grained early decisions get compressed, which is why re-pasting a hand-checked decision list helps.
Is /compact available in the claude.ai web app?
No. /compact, /clear, /context, and /rewind are Claude Code commands. On the web, do it manually: ask for a summary and paste it into a new chat.
How often should I start a new chat? For heavy, decision-dense work, every 15-20 messages or once the conversation passes roughly 60-70% of the window. Do not wait for the length-limit banner; quality slips before you hit it.
Related
- Claude usage limit
- Claude inaccurate answers
- Claude beginner guide
- Claude prompt best practices
- Claude Projects
Tags: #Claude #Debug #Troubleshooting