AI Guides › Claude Mastery
The Memory Feature: What It Remembers, What It Forgets, And Why That's Deliberate
By Nigel Guy · 2 min read
People react to Claude's memory feature one of two ways: they either ignore
it completely and re-explain themselves every session, or they assume it's
quietly logging everything they've ever said. Neither is accurate, and the
actual design is more deliberate than either extreme gives it credit for.
The rule: memory is opt-out by category and editable by you — it's a
working notebook it keeps for your benefit, not a surveillance log, and
treating it as either magic or menace means missing how to actually use it.
What it actually does
- It carries context across your chats, and — since it was unified in
2026 — across Claude Cowork sessions too. Something it learns in one place
carries to the other.
- It auto-extracts things like your role, your current projects, and your
stated preferences as you go. You don't have to fill in a form; ordinary
conversation is enough.
- Sensitive categories — health, religion, politics, ethnicity — are off by
default. They need an explicit opt-in from you before they're stored at
all.
- Certain categories are excluded outright, not just opt-in: things like ID
numbers and financial account numbers aren't kept regardless of settings.
- You can pause memory or fully reset it from Settings whenever you want. It
is not a one-way ratchet.
The mechanism for actually using it well
- Let it build naturally for the first week rather than trying to
front-load everything in one long message — see "The Five-Minute Setup"
for the initial pass, then let ordinary use fill in the rest.
- Check Settings once a month. See what it's actually retained; correct
or remove anything wrong or stale. This takes two minutes and prevents a
small early error from quietly shaping months of answers.
- Treat a correction as information, not an annoyance. If it recalls
something wrong, that's useful signal about what got mis-extracted — worth
a one-line fix, not a reason to distrust the whole feature.
- Reset deliberately when your situation changes. Changed jobs, wound
down a project, want a clean slate — a full reset is a legitimate,
supported action, not a workaround.
What to skip
Don't try to manually maintain a separate "memory document" alongside the
built-in feature — you'll end up keeping two things in sync for no benefit.
And don't assume that because something feels private, it's automatically
excluded; check the actual opt-in categories rather than guessing.
Guardrails
- Memory behaviour and its exact opt-in categories are the kind of detail
that gets refined over time — check your current Settings page rather than
trusting any guide's snapshot of it, this one included.
- Anything stored in memory should be treated as something you'd be fine with
if your account access were ever shared, audited, or reviewed at work.
Sensitivity settings are a control, not a guarantee of secrecy.
- Memory improves relevance, not correctness. A well-remembered preference
still doesn't excuse skipping the review step on anything that matters.
All 751 AI guides · JulieMango plans from £17/mo