Working memory and time
Combine leak, recurrence, adaptation and delay without freezing into a permanent attractor.
SENSORY RESEARCH
Research that includes the dead ends.
Sensory publishes mechanisms, test models, expected behavior and known limitations alongside results for reproducible technical discussion.
ACTIVE DIRECTIONS
State, attention and learnable structure.
Each direction connects an implemented mechanism to an open research question.
Dynamic competition
Test whether focus releases reliably and shifts to a relevant workspace competitor.
Fixed substrate, learned output
Separate stateful feature computation from a shallow supervised readout.
Learned conjunctions
Learn which multiplicative contextual features to create instead of designing them by hand.
OBSERVATIONS TO DATE
Implemented evidence, not universal claims.
Numerical results belong on the public website only with the exact model, dataset, seed, command and commit.
Normalization and adaptation
Limit runaway recurrence so a working trace can fade.
Oja and homeostasis
Preserve more distributed input weights than plain Hebbian concentration.
Isolated construct instances
Composition does not need to share hidden memory.
Shallow softmax
Currently works more reliably than deep CHL without weight symmetry.
KNOWN LIMITATIONS
The open problems are part of the result.
General deep credit assignment is unresolved, long-lived state can erode, gating remains mostly hand-designed and small experiments are not evidence of general intelligence or consciousness.