Claude Forgets Context in Long Chats: How to Fix

Why Claude Loses Memory in Long Conversations

When you have a long, ongoing conversation with Claude, you might notice it suddenly forgets earlier instructions, ignores previously established rules, or asks for information you already provided. This occurs because of how Large Language Models (LLMs) handle "context windows."

Every model has a maximum token limit (roughly equivalent to words and punctuation). While Claude has one of the largest context windows available, processing an extremely long chat history consumes massive amounts of memory. To keep performance fast and cost-effective, Claude eventually has to stop reading the very beginning of a long thread, or the system limits the active attention span of the model. When this threshold is crossed, Claude loses access to your initial prompts.

Here is how to troubleshoot and fix context memory loss in Claude.

1. Start a New Chat with a Status Summary

Continuing a single chat thread indefinitely is the most common cause of memory loss. Once a thread becomes too heavy, the oldest messages are effectively pushed out of the active memory window.

1. Copy the most important instructions, rules, and background details from your current chat.
2. Ask Claude in the current chat: "Summarize our progress, the established rules, and the current state of our project so far in a concise list."
3. Copy that summary.
4. Click Start New Chat in the top-left corner of the Claude interface.
5. Paste the summary and your primary instructions into the first prompt of the new chat. This resets the token usage to zero while preserving your progress.

2. Use Claude Projects to Anchor Key Information (Pro & Team Plans)

If you are a Claude Pro or Team user, you can use the "Projects" feature to keep critical information permanently in Claude's memory without it getting pushed out by chat history.

1. Navigate to the Claude home screen and click Projects in the left sidebar.
2. Create a new project for your task.
3. Under the Project Knowledge section, upload text files, PDFs, or documents containing your core instructions, codebases, or style guides.
4. Any chat you start within this project will automatically reference these files first. This keeps your foundational rules permanently accessible, even if individual chats grow long.
5. Use the Custom Instructions space within the project settings to define behavioral rules that Claude must always follow.

3. Structure Key Rules with XML Tags

Claude is trained to pay strict attention to data wrapped in XML tags. If you have specific rules that Claude must never forget, structuring them clearly helps the model's attention mechanism prioritize them.

1. At the beginning of your prompt, wrap your permanent rules in tags like <rules> or <instructions>. 2. Example: `xml <instructions> - Always write code in Python 3.10. - Do not use third-party libraries unless requested. - Keep explanations brief. </instructions> ` 3. If Claude starts drifting, you can quickly prompt: "Refer to the <instructions> tag in my first message and correct your last output."

4. Prune Your Uploads and Images

Images, screenshots, and large PDF attachments consume a massive number of tokens compared to plain text. If you upload multiple files in a single chat, you will hit Claude's context limits much faster, triggering memory degradation.

1. Avoid uploading the same document multiple times in a single thread.
2. If you must analyze a new version of a file, start a new chat rather than uploading it to the existing thread.
3. Convert images to text descriptions if possible, as image tokens consume a large portion of the active memory window.

When to Escalate

If Claude begins forgetting instructions after only two or three short messages, or if the interface frequently crashes and displays a "Failed to load chat" error, this points to an platform-level bug or browser caching issue rather than a standard context limit.

Check the official Anthropic status page to see if there is an active performance degradation incident. If the status is green, try clearing your browser's application data (local storage and cookies) or logging out and logging back in. If the issue persists across all new chats, contact Anthropic support through the in-app help widget.