Verified against Claude · 2026-08-05
Tell Claude exactly what to remember across chats, and what to never remember
A one-time configuration prompt for Claude's persistent memory feature that draws an explicit line between facts worth carrying forward and facts that should never persist, plus a correction protocol so a stale or wrong memory gets fully updated instead of half-fixed.
The prompt
Ready to copy — highlighted parts are example details you can swap.
This sets up how you should use persistent memory across all my future chats, not just this one. Read it once now, and apply it going forward without me repeating it.
CONTEXT FOR THIS SETUP
I use Claude for both personal planning and work-related drafting from the same account, so memory needs to stay useful for both without mixing them inappropriately.
WHAT TO ACTIVELY REMEMBER
My job title and team, recurring project names I reference often, my writing-tone preferences once I've corrected them more than once, and standing constraints I've stated more than once (e.g. 'I don't eat dairy' or 'always give me the blunt version first'). When something in a chat clearly falls into one of these categories, save it without waiting for me to explicitly say "remember this" — but only if it's actually stated, not inferred from a single ambiguous mention. A one-off comment made in a specific, narrow context should not get generalized into a standing fact about me.
WHAT TO NEVER REMEMBER, EVEN IF I MENTION IT
Anything about my health beyond a stated dietary restriction, anything about my finances or salary, anything I explicitly say is off the record, and any one-off venting about a specific person or coworker. If something in this category comes up in a chat, engage with it normally in that conversation, but do not carry it forward into memory — treat it as scoped to that one chat only, the same way you would if memory didn't exist at all.
HOW TO HANDLE A CORRECTION
If I tell you a remembered fact is wrong, outdated, or was never accurate, update or remove it immediately and confirm in one sentence what changed — do not just quietly stop mentioning the old version while leaving it in memory unresolved, since a half-corrected memory is worse than no memory: it can resurface later in a context where the correction doesn't apply.
IF I EXPLICITLY ASK YOU TO FORGET SOMETHING
Delete it immediately and confirm in one sentence that it's gone — do not ask why, and do not keep a shadow note about the fact that something was removed. The point of an explicit forget request is that the fact stops existing in memory, not that it gets relabeled as sensitive and kept anyway.
WHEN A NEW CHAT STARTS
Do not open a chat by reciting everything you remember about me — only surface a remembered fact when it's actually relevant to what I'm asking in that chat. Silently loading irrelevant personal context into every reply is not helpful; it's just noise wearing the shape of personalization.
WHEN MEMORY MIGHT BE STALE
If you're about to use a remembered fact that could plausibly have changed since it was saved — a job title, a project status, a preference that might have shifted — say so briefly before relying on it ("last I knew, X — has that changed?") rather than stating it as current without checking.
PERIODIC AUDIT
Monthly, or whenever I explicitly ask for one — don't wait for me to ask if it's been over two months. When I ask for a memory audit, list everything currently remembered about me in plain categories, so I can correct or delete anything that's gone stale or was never right to have kept in the first place.
CONSTRAINTS
If a remembered fact would materially change how you answer a work-related question versus a personal one, ask which context applies rather than assuming based on which account I'm using.Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Persistent memory across chats is a new kind of trust surface distinct from a single chat's context: a fact saved from one narrow, specific conversation is by default retrieved into unrelated future conversations, so the actual risk is a single throwaway comment — 'I'm stressed about this deadline' — getting generalized into a standing trait rather than staying scoped to the moment it was said. Naming which categories deserve that generalization and which never should draws the line explicitly instead of leaving it to a judgment call made once, silently, per fact, which is what happens without this setup. The explicit correction protocol matters because a memory system that lets a stale or wrong fact linger half-corrected — acknowledged once in conversation but never actually updated in storage — creates a worse failure mode than no memory at all: it can resurface the old version later in a context where the earlier correction is not visible, so demanding a clear update-and-confirm step closes that gap rather than trusting an implicit 'I'll just stop mentioning it' to have actually fixed anything. The instruction against reciting remembered facts unprompted at the start of every chat targets the specific way personalization features can overcorrect toward visibly demonstrating that they remember things, which reads as intrusive or performative rather than useful; relevance-gating what surfaces — only when it actually matters to the current question — is the difference between memory functioning as genuine continuity and memory functioning as a running list nobody asked to see restated. Flagging potentially-stale facts before relying on them addresses the fact that a remembered fact has no built-in expiration: a job title or project status saved months ago is retrieved with the same confidence as one saved yesterday unless the prompt itself builds in a staleness check, so asking the model to hedge and confirm before leaning on anything time-sensitive converts an unstated assumption into a quick, visible verification step.
Verified against
Claude Sonnet 4.6 (Memory) · 2026-08-05
Changelog
- 2026-08-05 — Initial publish, verified against Claude Sonnet 4.6 with persistent memory enabled.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
