Claude keeps forgetting earlier context — why it happens and how to fix it

Two distinct phenomena that feel like forgetting

• The first is a hard context limit: the conversation has exceeded 200,000 tokens and Claude tells you. This is rare in normal use.
• The second is attention drift: more common and more subtle. Claude, like all large language models, pays differential attention across a long context. In very long conversations, content near the beginning receives less attention than content near the end — so instructions established early can have less influence on later responses.

Signs of attention drift

• Claude ignores a writing style or format you established early on.
• It stops following a constraint you specified in your first message.
• Responses feel increasingly generic as the conversation progresses.
• It produces output that contradicts an earlier decision you both agreed on.

How to fix it

• Re-anchor key instructions mid-session — re-state critical constraints periodically rather than relying on a single mention at the start.
• Summarise and reset: ask for a mid-session summary and start a new conversation with it as the opening.
• Use Claude Projects for persistent instructions — these load fresh for every conversation within the Project.
• Put critical instructions at the end of a long conversation, not just the start — models attend more strongly to recent content.

Related guides

• Claude context window limit explained
• ChatGPT keeps forgetting context — how to fix it
• How to pick up where you left off in an AI chat

Frequently asked questions