How Claude Memory actually works — and how to use it

Claude Memory solves a specific problem: every Claude conversation, by default, starts from scratch. Claude does not know who you are, what you have worked on before, or what you care about. Memory changes that — it allows Claude to retain information across sessions, building up context over time rather than starting fresh every time.

Understanding how it actually works prevents the disappointment of expecting too much, and the underuse that comes from expecting too little.

What Memory stores and when

Memory is not automatic recall of everything Claude has ever seen. It is a set of short topic files that Claude writes and updates as you chat, rather than a summary generated after a conversation ends. You can read, edit or delete each one under Settings → Memory → Topics, which is the single most useful thing to know about it: when Claude has something wrong, you fix the file once and every future conversation inherits the correction.

By default Claude keeps sensitive subjects — health, beliefs, identity — out of memory. There is an "include sensitive topics in memory" setting if you want them retained. Some categories are always excluded regardless: identification numbers, criminal history, and anything that would violate the usage policy.

Things Memory is designed to hold:

• Facts about you: your role, preferences, how you like to work
• Facts about your organisation: company name, key context, ongoing projects
• Preferences about how Claude should behave with you: tone preferences, formatting preferences, things to avoid
• Ongoing context: a project you are working on, a decision you are tracking

Memory is scoped to your account, not shared across a team — your colleagues do not inherit your preferences. It does follow you between surfaces, though: the memory you build in chat is the same memory Claude uses in Cowork, so a task running in the background already knows what you established in conversation.

How to build useful Memory

Memory is most useful when you treat it deliberately rather than hoping Claude automatically retains the right things.

At the start of a relationship with Claude, you can explicitly ask it to remember key context:

• "Remember that I am a head of product at a 50-person SaaS company."
• "Remember that I prefer bullet-point summaries over prose paragraphs."
• "Remember that when I ask for a draft, I want three options, not one."

These instructions persist. Future conversations start with Claude already knowing these things.

You can also just look. Settings → Memory → Topics lists everything, and reading it is faster and more reliable than asking Claude what it knows — the files are short, and seeing them is how you notice the one that has been quietly wrong for a month.

What Memory does not solve

Memory is not a knowledge base. Memory stores facts and preferences, not documents. If you want Claude to reference your company's product documentation, that belongs in a Project's uploaded files — not Memory. Memory is for persistent context about you and how you work; Projects are for domain knowledge about your work.

Memory is not a full transcript. Claude writes topics as you chat, so more gets captured than under the old explicitly-stored model — but it is still a set of distilled notes, not a recording. If you had a detailed strategic discussion last week, expect Claude to retain the shape of it, not the specifics. Searching your past chats is a separate capability from memory; when you need the actual content of an earlier conversation, that is the tool.

Memory goes stale, and stale memory is worse than none. Claude will confidently apply a fact that stopped being true in March. You can correct it in conversation — "forget that I'm head of product, I've changed roles" — but editing the topic file directly under Settings is more reliable, because you can see what actually changed. Keeping memory accurate is ongoing work, not one-time setup.

Memory for individuals vs. teams

Memory is fundamentally an individual feature. It personalises Claude for one person's way of working.

For team consistency — everyone getting the same baseline context — Projects with well-configured system prompts are the right tool. Memory adds personalisation on top of that: a CS rep might have individual Memory that tells Claude their preferred communication style, layered on top of a CS Project that gives everyone the same product knowledge.

Think of it as: Project = team context, Memory = individual personalisation.

The practical pattern

Set up Memory early with the four or five facts about yourself and your work that would most change how Claude interacts with you. Then actually open Settings → Memory → Topics once a quarter and read what has accumulated. Use it primarily for how you like to work, not for what you are working on (projects come and go — Memory is for the more persistent stuff).

For teams, do not rely on Memory for consistency. That is Projects' job. Memory is the layer that makes Claude feel like it knows you — useful, but not the foundation of team deployment.

Further reading

• Memory tool documentation — how memory works in the API, including the data model and persistence behaviour
• Claude's memory works everywhere, and you decide what's in it — the current model: topic files, shared across chat and Cowork, editable in Settings
• Use Claude's chat search and memory — memory and past-chat search, and how they differ