Three features, and one question that tells you which to reach for. The question is not which box does this go in. It's how widely should this apply?
Lesson 2 left you with an accounting habit: for any response, what information was actually available, and how did each piece get there? One honest answer is because I typed it. That works. It also means the brief you retype every Monday morning, the house style you paste into every chat, and the three sentences of background you have written out forty times are all costs you keep paying. This lesson is about the machinery that pays them once.
Almost every explanation of this — including an earlier revision of this course — presents three tidy containers: Instructions for Claude for global things, Projects for project things, memory for remembered things. It's a clean picture and it is wrong in a specific way that costs you later, so it's worth correcting before you build the habit rather than after.
The three are not parallel. Memory is not a fourth kind of thing sitting alongside Projects — each project has its own separate memory space and its own project summary, kept apart from other projects and from your non-project chats. source So memory has a project scope. Chat search adds a third retrieval route with boundaries of its own. Learn them as boxes and you will eventually watch the boxes leak into each other and conclude the whole thing is unpredictable.
Learn them as answers to one question and nothing surprises you:
| How widely should this apply? | Where it lives | Who puts it there |
|---|---|---|
| Everywhere, in everything I do | Instructions for Claude — Settings → Instructions for Claude. Account-wide, always on, applies to every chat including chats inside projects. | You, deliberately, once. |
| To this one body of work | A Project — project instructions plus project knowledge, available to every chat in that project and to nothing outside it. | You, deliberately, per project. |
| Whatever scope I happen to be working in | Memory — Settings → Memory → Generate memory from chats. Builds itself from your conversations, with a separate space per project. | Claude, as you go. You edit. |
Account-wide and always on. Everything you put here applies to every chat you ever open, including chats inside every project, forever, until you change it. That's the whole feature.
Which makes the bar high, and most people set it far too low. The test is not "would this be useful?" — nearly everything is useful. The test is would this still be true if you changed employer tomorrow, and is it still true on a Sunday? Very little passes.
| Passes | “I'm a solicitor. Don't caveat legal points into uselessness — I'll do the hedging myself.” “British English, always.” “Answer first, reasoning after.” “Metric units.” “I read faster than I listen; keep prose tight and skip the recap of what I just asked.” |
| Fails | “Our fiscal year starts in April.” “This client hates bullet points.” “Use the Lumen voice guide.” These are true of a body of work, not of you. Put them account-wide and they are in force in every chat where they don't hold. |
It's about you: your role, your register, how you want to be spoken to. If you can't finish the sentence "this is true of me whatever I'm working on," it belongs one scope down.
A project holds two things, and people reliably use one and skip the other.
Project knowledge is the material: the style guide, the past examples, the spec, the transcript, the price list. You add it from the knowledge panel on the project's page. Project instructions are the standing brief: who you are in this work, what you want done with that material, and what shape the output takes. Both reach every chat you start inside the project, without you retyping either.
The common failure is using a project as a folder — six files uploaded, no instructions written — and then typing the brief into the first message of every chat anyway. That's the whole cost you came here to stop paying, reintroduced one message at a time.
The worry people bring to project knowledge is that a big knowledge base gets truncated or crowds out the conversation. Large project knowledge is handled by retrieval — the relevant parts are found and brought in for the response that needs them, rather than everything being held at once. So the question you should actually be asking about a file is not "is this too much?" but "is this the material a good answer would need to draw on?" Curate for relevance, not for size.
One boundary worth knowing now, because it's the same boundary in both directions: from inside a project, searching your past chats covers that project only. From a chat outside all projects, it covers your non-project chats and reaches into no project at all. Project scope is a real wall, not a soft preference.
The other two are things you write. Memory is the one that builds itself: turn it on at Settings → Memory → Generate memory from chats and Claude accumulates individual entries, organised into categories, updated as you go rather than on some nightly schedule. Your role, your working preferences, details of things you keep coming back to.
The important structural fact, again, because it's the one that breaks the box model: memory is not a single account-wide store. Each project gets its own memory space and its own project summary, separate from your other projects and from your non-project chats. Work in a project and memory accumulates at that project's scope. Work outside one and it accumulates outside.
A lot of people are quietly uncomfortable with memory and can't name why, so they leave it off and never revisit the decision. That's a worse outcome than either using it or refusing it on purpose. Here is what actually exists, stated as fact rather than as reassurance — the controls are the reassurance, and you should judge them yourself.
| Read it | Every entry is a listed statement you can open and read. Not a vague cloud of recollection — discrete lines of text. |
| Edit it | Entries can be changed or removed individually. |
| Pause it | Keeps everything already stored and stops new entries being created. Not the same as reset. |
| Reset it | Permanently deletes all memories, including per-project ones. Irreversible. |
| Export it | Take the store out as a file. Import works too, which is the same fact viewed from the other end: it is portable, not locked in. |
| Bypass it | An incognito chat is excluded from memory and from chat search entirely. Use it for the conversation you'd rather didn't shape the next hundred. |
| Separate it | Per-project spaces mean the client work doesn't bleed into the job hunt. |
And the fact that is not reassuring, which is exactly why it's here: deleting a conversation does not remove the memory entries that came from it. Entries persist until you remove them yourself. source If your model of memory was "it fades when the chat goes", correct it now. Deleting the chat and clearing the entry are two separate actions.
This is the single most common complaint from people who use Claude every day, and underneath it is one conflation: said in a conversation and stored at a scope get filed as the same event. Both feel like telling it. Only one of them persists.
Once everything is “I told it”, the behaviour looks arbitrary — sometimes it knows, sometimes it doesn't, no pattern. And people don't usually escalate that into a complaint. They quietly downgrade how much they trust the tool, which is a much harder thing to recover from.
The repair is not conceptual. It's five minutes with the settings open. Go to Settings → Memory and read your own entries. Nothing else lands as hard as seeing discrete, listed, editable statements in the place where you had imagined something vaguer. It's in the exercise below, and it's the step to do even if you skip the rest.
Three places, so you recognise them rather than being surprised by them.
Memory inside a project isn't the same memory as outside it. Something Claude picked up while you were working in the Lumen project stays at that project's scope. Open a chat outside every project and expect it there and you'll be disappointed, and it will feel like forgetting. It isn't. It's scope.
Chat search has a wall in the middle. Outside projects it covers your non-project chats. Inside a project it covers that project. There is no view that spans both, so "I definitely discussed this with Claude" is only half of a location.
Instructions for Claude apply inside projects too. This is the leak that runs the other way, and it's the one that catches people out: a global instruction you wrote nine months ago is still in force in the project you set up this morning. If a project's output has a persistent oddity you can't trace to the project instructions, that's the first place to look.
What to do about it: when something doesn't carry across, don't ask whether Claude remembered. Ask which scope did I store that at, and which scope am I standing in now? Those two questions resolve nearly every instance of this.
One build, five steps, about twenty minutes. Nothing here has a required outcome — step 3 asks you to compare two runs, and "they came out much the same" is a result, not a failure. Write down what you expect before each run, then what actually happened.
You need two files, both supplied, neither of them from your job:
Thirteen rules with real teeth, a newsletter format spec, a revision history, and one argument nobody has won.
File 2Past issuesIssues 41, 44 and 47, as sent. What the guide produced in practice, which is not quite what it says.
Plain-text versions for uploading: lumen-brand-voice.txt and lumen-newsletters.txt. Save both before you start. Lumen & Co. is fictional and turns up again in Lesson 7, so a project you build now is worth keeping.
1 · Build the project. Create a project called Lumen newsletter. Add both files to the knowledge panel. Then write the project instructions — and they must include at least one output-format constraint, the kind you learned to specify in Lesson 3. The material hands you several: three or four sections, each 60–120 words, every section ending with exactly one link whose text is a verb phrase, sign-off as a first name and a job. Pick one, or more. Write them as instructions, not as a description of the guide.
2 · Run the task in the project. New chat inside the project. Ask for the next issue's section on one of these: the Waxed Field Jacket coming back into stock, a price rise on mending, or the Quilted Gilet selling out. Don't paste the guide. Don't describe the voice. Just ask for the section.
3 · Run the identical task in a bare chat. Same words, outside every project, nothing attached. Put the two outputs side by side and write down three concrete differences — not "better", but what differs. Length. Whether the links are verb phrases. Whether a number appears where the guide demands one. Whether the banned words show up. That delta is the argument for Projects, and it holds whichever way the two runs come out.
4 · Find something that deserves the widest scope. Write one line of Instructions for Claude that would be true of every project you would ever run — this one, your actual job, whatever you do at the weekend. Then write down how many candidates you rejected getting there. The count is the lesson.
5 · Read your own memory. Open Settings → Memory. If it's off, this is a good moment to decide on purpose rather than by default; if it's on, read the entries. Pick one and write a sentence saying why it belongs at the scope it's at — and whether you'd have put it there. You are not asked to delete anything. Inspection is the whole point of this step.
Five boxes. Tap as you go — it saves in this browser only.
The five steps are the exercise; the newsletter is just a vehicle. Two alternatives that work identically:
Weekly status roll-up. Knowledge: a document describing what your weekly update has to contain, plus two past updates. Format constraint: a fixed section order, or a hard cap on lines per workstream.
Candidate screening notes. Knowledge: a role description plus your criteria, and two past write-ups. Format constraint: one paragraph per criterion, a recommendation on its own line, no adjectives about the person.
If you use your own material, keep it to things you'd be comfortable having sit in a project. The Lumen files exist precisely so that nothing in this lesson requires you to upload anything real.
If you want to practise editing and deleting memory entries, make one first. In an ordinary chat, state a deliberately disposable and obviously false preference — something like "for the record, I always want dates written in Roman numerals" — carry on for a couple of turns, then check whether an entry appears at Settings → Memory. Edit it. Then delete it. You have now used both controls without touching anything you actually rely on. Note whether the entry appeared at all, and how it was worded compared with what you said.
From memory, without scrolling up: name the three scopes and the one question that routes between them. Then name one fact about memory that is true whether or not it makes you comfortable.
The question: how widely should this apply? The scopes: everywhere → Instructions for Claude (account-wide, always on, about you); this body of work → that Project (instructions plus knowledge, and nothing outside sees it); accumulating on its own → memory, at whichever scope you're working in.
Facts about memory that hold either way: each project has its own memory space and project summary; entries are individual, listed and editable; you can pause, reset, export and bypass it with an incognito chat; and deleting a conversation does not remove the entries that came from it.
Not which box — how widely should this apply?
True about me everywhere → Instructions for Claude.
True about this body of work → that Project.
Happy for it to accumulate → memory, at the scope you're in.
Anthropic, Use Claude's chat search and memory to build on previous context — the help article this lesson's factual claims were checked against. support.claude.com/en/articles/11817273
Read two things and you can skip the rest. First, the section on what chat search covers — it states the project boundary explicitly, and that sentence is the whole of this lesson's "where the boxes leak". Second, the controls list, so the pause/reset distinction comes from the source rather than from me. Skip the setup walkthrough; you'll do that faster by opening the settings.