I came back to a client site's copy after two weeks away. New chat, simple request: pick up where we left off. Claude started writing without hesitation.
Then I read it back and stopped. The whole piece had been written around a structure I had decided, two weeks earlier, to abandon.
My first assumption was that I had asked badly. But rephrasing landed me in the same place every time. The problem was not my instructions. It was that I had been treating two separate mechanisms as one.
Claude has two routes back into your earlier conversations: memory, and search over past chats. They look similar from the outside, but they store things in different places, reach different distances, and are cleared in different ways. Once I stopped conflating them, "the continuation" finally continued from where I actually was.
One remembers, the other goes looking
Memory works as Claude chats. It writes down individual topics in the moment, rather than summarizing a conversation once it ends. You can read everything it holds under Topics in your memory settings, open any entry, edit it, or delete it.
Past-chat search works on request. Ask Claude to find what you discussed before and it searches your saved conversations and brings back only the relevant slice. The help center describes this as retrieval-augmented generation, and on screen it shows up as a tool call.
| Aspect | Memory | Past-chat search |
|---|---|---|
| When it acts | Saves during the chat; already in effect when the next one opens | Only when you ask, and only at that moment |
| How you inspect it | Listed as Topics in memory settings | Follow the citations in the reply back to the original chat |
| What it is good at | Your role, ongoing work, writing preferences — the slow-moving background | What you decided on a particular day; exact wording and numbers |
| Where it fails | Stale entries linger, and the conversation proceeds on an old premise | Without a phrase to search for, nothing surfaces |
| Plan availability | On by default for Free, Pro and Max (owners enable it for Team and Enterprise) | Paid plans only: Pro, Max, Team, Enterprise |
That last row is the one worth sitting with. On the free plan you get memory but not search. So "Claude clearly remembers me, yet it cannot retrieve that one conversation" is not moodiness — it is the plan boundary.
Who you are goes in memory; what happened goes in search. Drawing that line removed most of my hesitation about how to ask. My habits and my preferred heading shapes can live in memory quite happily. But "which structure did I settle on two weeks ago" is an event, and events are safer retrieved than recalled.
The citation tells you which one answered
When a search runs, a tool call appears on screen and the reply carries citations back to the original chats. When Claude is speaking from memory, there are no citations. So the presence or absence of a citation tells you what the answer is standing on.
On the day I got that outdated draft, no tool call had appeared at all. Claude never went looking — it wrote from the stale premise sitting in memory. The evidence had been on screen the whole time.
For work I have been continuing over weeks, I no longer ask for the continuation directly. I put one preface in front of it:
Before writing the continuation, split this into two parts for me.
1. What you currently hold in memory as the premises of this project
2. What you find by searching past chats as the most recent decisions
For (2), show me the chats you are citing.Seeing the premises and the decisions side by side surfaces the mismatch before a single paragraph gets written. If a stale premise shows up, I say so in the chat — "that premise has changed" — and the correction lands in memory for the next conversation. Being able to fix it without opening settings is one of the kinder parts of this feature.
Projects do not reach across their own walls
The help center gives search exactly two territories: all chats that live outside projects, and the conversations inside an individual project. The second is bounded by that project alone.
So one project cannot search another. A conversation you had outside of projects will not reach you from inside one. Memory follows the same shape — each project keeps its own separate memory space and its own summary.
As an indie developer juggling client work, I had been scattering drafts between a project and the open chat list depending on my mood that day. This explained, at last, why some conversations simply refused to be found. One job now lives in one project, without exceptions. I wrote separately about managing context with the Projects feature if that side is what you are sorting out.
Cowork has a matching boundary. Memory is shared between chat and Cowork only when Cowork runs in the cloud; sessions running locally on your own machine do not use memory at all. If the context from a local task never shows up in your next chat, that is the design, not a fault.
Deleting a chat does not delete the memory it created
This is the part I see missed most often. When a conversation is deleted, or simply expires, the memory entries generated from it are not removed. To clear them you go into memory settings and delete the topics one at a time.
The legacy experience behaved the opposite way: deleted conversations dropped out of the memory synthesis, which was rebuilt every 24 hours. So if you are running on the old mental model — "I deleted that chat, so the premise is gone too" — the premise is in fact still sitting there untouched.
You can tell which experience you have from where the setting lives. A Memory entry in settings means the newer experience; memory tucked under Capabilities means legacy. The help center noted that legacy memory could be exported until September 9, 2026, so if you are looking at the newer screen today, you are on the far side of that migration. For background on when memory first opened up broadly, see the article from when memory rolled out to all users.
My monthly pass takes three steps:
- Open memory settings and read the Topics from the top. Role, ongoing work, preferences — those three categories are enough
- Rewrite entries whose facts have changed, rather than deleting them. Leaving a correct premise in place beats leaving a gap
- Delete entries for work that has finished
Five minutes, once a month. I do it on the first of the month.
Keep throwaway thinking somewhere that does not persist
The ghost icon in the top right of a new chat opens an incognito chat. Nothing there is saved to your history, nothing enters memory, and nothing becomes searchable later. It is available on every plan, free included.
I use it for spreading out options I might well throw away. Two of three ideas never become anything — and if that conversation happens in a normal chat, the discarded ideas settle in as premises, and I spend time later walking them back. Not creating the residue is far lighter than clearing it.
Team and Enterprise plans work differently here, though. Incognito chats are still included in organizational data exports and follow your retention policy. Disappearing from your own history and disappearing from the organization's records are two separate things, and it is worth knowing which one you are getting.
The single thing worth checking today is the Topics list in your memory settings. Five minutes will cover it. What is written there is what speaks on your behalf, before you do, the next time you open a chat.
Since the day I reread that draft built on a two-week-old premise, that list has been part of my first-of-the-month routine. The longer a piece of work runs, the more it pays to have decided what Claude should remember and what it should go and find.
Thanks for staying with a piece about settings screens — I know that is not the most thrilling subject to give your afternoon to.