Reference-Back
Paradigm

Working-memory gating

The n-back task asks you to update working memory on every single trial, which makes it impossible to say what part of your score is updating and what part is everything else. The reference-back task fixes that by making updating optional, trial by trial — so the cost of opening the gate to memory, closing it again, and updating what is inside can each be measured on their own.

The task

A letter — X or O — appears inside a coloured frame. You always answer the same question:

Is this letter the same as the letter in the most recent red frame?

X
Red frame. Answer as usual — then this letter becomes the new one to remember. It replaces whatever you were holding.
O
Blue frame. Answer as usual — and keep holding the same letter. A blue letter never becomes the thing you remember, however many of them go by.

Press L or → for same, A or ← for different. On a phone, tap the right half of the screen for same and the left half for different.

Worked example

X
—
remember X
O
different
vs X
O
different
vs X, now hold O
X
different
vs O, now hold X
O
different
vs X
X
same
vs X
X
same
vs X, still X

Trial 6 is the one that matters: the answer is same because you are still holding the X from trial 4 — the blue O on trial 5 went past without touching it. Getting that trial right is the whole task.

Session length

The frame colour changes on only 25% of trials, and those are the trials the gate costs are built from — so they, not the total, decide how steady the numbers are. These options give 16, 28, 42 trials per switch condition.

Worth knowing before you pick: these costs sit on top of reaction times that vary by more than a hundred milliseconds, so a short session can leave a real cost indistinguishable from zero. Simulating a participant carrying the published effect sizes through this implementation, gate closing separates from zero in 90% / 99% / 100% of sessions at the three lengths, gate opening in 48% / 72% / 89%, and substitution — a double difference, so much the noisiest — in only 26% / 39% / 52%. The report greys out any cost it can't resolve rather than dressing it up. That is not wasted effort: unresolved sessions still feed your trend, where averaging across sessions recovers what no single one can show.

Advanced parameters

You have the whole stimulus-plus-interval window to answer, so these two together set the response deadline. The switch rate is deliberately not adjustable: changing how often the colour flips changes how strongly you expect it to stay, and that moves the switch costs for reasons that have nothing to do with your gating.

Each block opens with a red letter to memorise. Blocks are separated by a break you control.

Method follows Rac-Lubashevsky & Kessler (2016, Neuropsychologia 90, 190–199): the same 25% frame-colour switch rate, X/O letters, and the same four orthogonal contrasts — gate opening (reference-switch − reference-repeat), gate closing (comparison-switch − comparison-repeat), updating (all reference − all comparison) and substitution (the trial-type × match interaction on repeat trials). It differs in showing each letter for a fixed interval rather than until you respond, and in running a few hundred trials rather than their 900. This is a research-style instrument and educational demonstration — not a medical or diagnostic device. A single session is noisy and the overlap between populations is large.

Working-memory gating

Reference-Back Paradigm

What this test measures

The reference-back paradigm (Rac-Lubashevsky & Kessler, 2016) refines the classic n-back to separate processes the n-back confounds. Letters appear in red or blue frames, and you always make the same judgment: is this letter the same as the one in the most recent red frame? Red frames update your stored reference; blue frames must be judged against it without overwriting it.

Because the response demand is identical on both frame colours while the memory demand differs, the task can isolate the components of working-memory gating: the cost of opening the gate to let new information in, the cost of closing it to protect what is stored, and the cost of the update itself.

How to use it

The judgment is simple but the gating costs are subtle, so consistent, attentive runs matter more here than in most tasks. The task flags runs whose estimates look unreliable. Signed-in sessions feed the working-memory-gating trend on the Progress page.

Sapiens Bench tasks are educational instruments for self-tracking, not medical or diagnostic devices.