Memory across chats: what persists and what does not

The model itself remembers nothing between requests. When a new conversation seems to know something about you, a separate feature wrote that something down and pasted it back in. Memory is a store, and the store is yours to read.

Getting this distinction straight explains most of what people find eerie or unreliable about it.

Topics, not transcripts

Claude saves memory as a set of individual topics while you chat. It does not wait until a conversation ends and summarise it. You can also say "remember this" and have it saved directly.

That carries a practical consequence: what persists is a claim about you, extracted and stored, and an extracted claim can be wrong in a way a transcript cannot. A passing remark can be recorded as a settled preference. An example you invented to illustrate a point can be filed as a fact about your situation.

It is a list you can open

Everything Claude has stored appears under Topics in Memory settings, where each entry can be read, edited or deleted. You can also ask for changes in the conversation itself.

Most people never look. Doing it once is genuinely clarifying, and it is the only way to find the entry that has been quietly skewing answers for a month.

Projects are walled off

Each project keeps its own memory space and its own project summary, separate from other projects and from your ordinary chats. Work you do inside a client's project does not leak into the next one, and context stays scoped to the thing it belongs to.

This is the main argument for putting work in a project rather than starting another standalone chat.

What it declines to store

By default Claude does not store topics touching personal or sensitive subjects: health, race, ethnicity, religious beliefs, politics, gender identity and similar areas. A narrower set is never saved at all — government ID numbers, criminal history, financial account numbers and immigration status.

Read that as a sensible floor. It is no guarantee that it covers everything you would consider private. It is a defined list, and your own sense of what is sensitive will be wider than any list someone else wrote. The rules in what never goes in a prompt still apply.

Try this

Open Memory settings and read every topic stored about you. For each one, ask whether you would have said it aloud in those words. Delete the entries that are stale, over-generalised, or were true for one task and are being applied to all of them.

What goes wrong

Assuming it remembers the conversation. It remembers extracted topics. The reasoning, the caveats and the specific numbers from last week's chat are gone unless a topic captured them, and a topic is one sentence long.

Letting a wrong entry compound. An incorrect stored preference gets applied silently to every new conversation, and each answer it shapes looks like a fresh judgement. This is the failure mode that costs the most, because nothing about the output points at its cause.

Confusing it with the context window. Memory carries a little between conversations. The context window is what the model can see right now, and it is emptied every time you start again. Two mechanisms, two failure modes, and people routinely diagnose one and fix the other.

Treating a project summary as a project brief. It summarises what has happened, so it drifts toward whatever you have talked about most, which is seldom whatever matters most. Put the things that must stay true in project knowledge, where you control the wording.

How to check it worked

After editing your topics, open a new conversation and ask: "What do you already know about me and my work?" Compare the answer against the list you just curated. Anything that appears and should not is an entry you missed; anything missing that you expected tells you the store is thinner than you assumed, which is useful before you rely on it.