Why Claude Loses Context (and 4 Fixes)
You spent twenty minutes setting up the constraints. Claude agreed to all of them. Now it is proposing the exact approach you ruled out at the start, and you are not sure whether to correct it again or start over.
Or you open Claude the next morning, and the project you worked on yesterday might as well have never happened.
Those are two different problems with two different fixes, and most advice online treats them as one. Below are the four reasons Claude actually loses context, how to tell which one you are hitting, and what to do about each.
First, work out which failure you have
Ask yourself one question: did Claude forget something from this conversation, or from a different one?
Forgetting inside the current thread is a context window problem. Causes 1 and 2.
Forgetting across threads is a memory problem. Causes 3 and 4.
They feel identical when it happens. They have nothing to do with each other.
Cause 1: the conversation hit its maximum length
Claude reads the conversation by loading it into a context window. The window is large, but it is finite. When a thread gets long enough you will see it stated plainly: Claude hit the maximum length for this conversation.
At that point Claude is not being difficult. It physically cannot read the whole thread anymore.
The fix: do not try to rescue the thread. Start a new one and bring the conclusions with you.
Before you leave the old conversation, ask Claude:
Summarise the decisions we made in this conversation. List the constraints I gave you, the approach we settled on, and anything I explicitly ruled out. Keep it under 300 words.
Paste that into the first message of the new chat. A summary you have read and approved is worth more than forty messages Claude can no longer see.
When it will happen again: any thread where you paste large documents, long code, or many rounds of revision. The window fills faster than most people expect.
Cause 2: the thread is still open, but the early instructions have faded
This one is more confusing, because nothing has visibly broken. The conversation is still going. Claude just quietly stops honouring something you said at the beginning.
Even inside the window, the model does not weight every message equally. Recent turns dominate. Instructions given forty messages ago are still technically present but carry less influence than what you said two messages ago.
The fix: restate the constraint, and restate it late.
Rather than pointing back to it, say it again as if it were new:
Reminder of the constraints for this task: no external dependencies, output in TypeScript, keep functions under 20 lines.
It feels redundant. It works better than a reference to something buried upstream.
A cheaper habit: for any long working session, keep your constraints in a short block you can paste again every ten or so messages. Rewriting it takes seconds. Re-doing the work Claude did wrong takes longer.
Cause 3: you are in a Project and expecting it to know everything else
Projects hold their own knowledge and their own conversations. That is the point of them. It also means a Project is a boundary.
Work you did in a different Project, or in a normal chat outside Projects, is not automatically available inside this one.
The fix: put the durable material into the Project's knowledge rather than into a single conversation inside it. Anything you want available in every chat in that Project should live at the Project level, not in one thread.
If you are switching between Projects and expecting continuity, that continuity does not exist. Move the material deliberately.
Cause 4: memory is on, but it stored less than you think
Claude memory has been available on Free, Pro and Max plans since March 2026, and the relevant settings are on by default. So the usual problem is not that memory is switched off. It is that what Claude kept is thinner than what you remember telling it.
The fix: go and look, do not guess.
Open Settings, then Capabilities, then Memory. Two switches sit there, Search and reference chats, and Generate memory from chat history, normally both on. Below them, View and manage memory shows what has actually been stored, grouped by category.
Read that list rather than assuming. Asking Claude directly is faster still:
What do you currently remember about me and my work?
The answer is usually shorter than expected. If something important is missing, it was never stored, and no amount of restarting the conversation will bring it back.
Two cases where the behaviour genuinely differs. Team and Enterprise accounts are still on the older summary-based memory while the newer experience rolls out, so guides written for personal plans may not match what you see. And searching past conversations on demand is a paid feature, so on a Free plan Claude can carry forward what it already stored without being able to go back and look up the specific chat you have in mind.
Which fix applies to you
If you saw the message about hitting the maximum length for this conversation, that is Cause 1. Summarise the decisions and start a new thread.
If the thread is still open but early rules are quietly being ignored, that is Cause 2. Restate the constraint late, rather than pointing back to it.
If you are inside a Project and expecting knowledge from somewhere else, that is Cause 3. Move the material into the Project's knowledge.
If it is a new chat and Claude knows less about you than you expected, that is Cause 4. Open the memory list and read what is actually there.
If you cannot place your symptom in any of those, it is worth checking whether the thing you expected Claude to remember was ever stored in the first place.
The one thing none of these fix
Every fix above keeps your context inside Claude. None of them move it.
Open ChatGPT and it does not know what you and Claude worked out this morning. Open Gemini and it starts from nothing. If Claude is the only assistant you use, this is irrelevant. If you use several, it is the version of the problem you hit most often, and no Claude setting addresses it, because the memory belongs to the platform rather than to you.
There are three practical ways to handle it, and they are worth comparing honestly.
Keep a document. A running notes file with your standing context, pasted into the first message of important chats. Free, works everywhere, and most people abandon it within a fortnight.
Export and re-import. Some assistants can now pull memory across from another provider. Fine for a one-time move, too slow to be a habit.
Put the memory outside the assistants. Keep one context layer of your own and let each assistant read from it. This is what Unibase Memory does. It captures context from your chats and lets you send the relevant parts into ChatGPT, Claude, Gemini, Grok, Kimi, or DeepSeek, so the same background follows you instead of being retyped.
It is a Chrome extension, free to start, and it does nothing for Causes 1 through 4 above. It solves a different problem: the one where Claude's memory works exactly as designed and simply cannot leave.
If you want the manual version first, we wrote out the step-by-step process for moving context between assistants, including what each platform lets you export. If ChatGPT is the one giving you trouble, the causes there are different enough to need their own guide, and the same is true of Gemini.