uygnoey Memory curation
Claude's memory is shared between chat and Claude Cowork: one store, visible as a list of short topic files, editable by the user. This skill covers reading that store, correcting it, and deciding what should be allowed into it.
The behavior described here comes from the product announcement of 2026-08-25; see references/memory-controls.md for the settings, storage rules, and availability details in full.
Instructions
1. Establish what is actually stored
Memory content lives in Memory settings under Topics, as a list of files. Before advising on anything, have the user open that list rather than guessing:
- Each entry is a short file that can be read, edited, or deleted individually.
- The same store is used by chat and by Cowork, including Cowork tasks running in the cloud.
- Claude adds topics during a conversation, not by summarizing it afterwards, so recent turns may already be reflected.
If the user cannot find the list, the path is Settings > Memory.
2. Correct at the source, once
When Claude uses stale or wrong context, the fix belongs in the topic file, not in a per-conversation correction:
- Locate the topic file holding the wrong fact.
- Edit that one file (for example, replace a company's old name with the current one).
- Every conversation from then on picks up the corrected version — in chat and in Cowork.
Prefer editing over deleting when only part of a file is wrong; delete the file when the whole topic is obsolete. Walk through templates/topic-review-checklist.md for a repeatable pass over a single file.
3. Use memory instead of re-briefing
If the user is repeating context across surfaces, the answer is usually to explain it once in chat and let it carry:
- Say who a recipient is and how they like updates written, then ask Cowork for the draft.
- Work out a plan in chat (headcount, city, speakers), then ask Cowork for the doc built on it.
- Define the team's metrics once; later decks reuse those definitions without rebriefing.
Worked situations from the announcement are collected in examples/curation-scenarios.md.
4. Scope sensitive topics deliberately
By default Claude does not store personal or sensitive subject matter — health, race, ethnicity, religious beliefs, politics, gender identity, and similar areas.
When a user asks Claude to remember something in one of those areas:
- Explain that the "include sensitive topics in memory" setting governs it.
- State the three properties that matter for the decision: saves are announced with a notice, storage applies going forward only (nothing before activation is captured retroactively), and the setting can be turned off at any time.
- Let the user make the call. Do not push in either direction — what one person considers sensitive, another considers useful context.
5. Know the hard limits
Some categories are never stored, even with sensitive topics turned on:
- Sensitive identification numbers (SSN, government ID numbers)
- Criminal history
- Immigration status
- Anything that violates the Acceptable Use Policy
Claude informs the user when it cannot update memory with such information. Do not offer workarounds for these categories; note the limit and move on.
6. Check availability before troubleshooting
If memory appears to be missing entirely, check the plan before anything else:
- Free, Pro, Max — memory is on by default across web, desktop, and mobile; saving sensitive topics is off by default. On iOS and Android, the app must be on the latest version.
- Team, Enterprise — admins control availability for the organization, and memory is off for each individual user until they turn it on.
7. Pause or reset when the context is wrong for the work
Memory can be paused or reset at any time. Suggest pausing for a session that should not influence later ones, and resetting when the accumulated context no longer matches the user's situation. Both are reversible choices the user makes; make the consequence explicit before a reset, since it clears what has been built up.
Examples
"What does Claude actually know about me?" Point to Settings > Memory > Topics. Explain that each entry is a short, readable file, that the same list backs both chat and Cowork, and that anything there can be edited or deleted.
"Claude keeps calling my company by its old name." This is a stored-fact problem, not a prompting problem. Have the user open the topic file containing the company name and correct it there. One edit fixes every later conversation on both surfaces instead of a correction repeated per chat.
"I told Cowork about my Q3 priorities last month and it forgot." Check availability first: on Team or Enterprise, an admin must enable memory and the user must turn it on individually. If memory is on, check the Topics list — if the priorities are not there, they were never captured; restate them in chat and confirm the topic appears.
"Remember that I'm gluten intolerant so meal plans work." Health is a sensitive topic and is excluded by default. Explain the "include sensitive topics in memory" setting, that saves in these areas come with a notice, that it only applies going forward, and that it can be switched off later. Then let the user decide.
"Save my passport number so I don't have to look it up." Sensitive identification numbers are never stored, regardless of the sensitive-topics setting. Say so plainly, note that Claude reports when it cannot update memory with such information, and do not propose an alternative storage trick.
"Our team's QBR decks keep using the wrong definition of active users." Have the user state the team's metric definitions once in chat and confirm a topic file was written for them. Later Cowork decks then build on those definitions with no rebriefing; if a definition changes, edit the file rather than restating it per deck.