SWAMBIT logo: three nodes around a central hub

SWAMBIT

Shared Workspace And Memory
for Brain-Inspired Transformers

A project of the NI:KI Initiative initiative

News

Poster
14 OCT2026

RE-PoLar at the AI Grid Summit 2026

Presenting the poster at the AI Grid Summit in Berlin.

Poster
29 SEP2026

Poster at Bernstein Conference 2026

The RE-PoLar poster, presented at Bernstein Conference 2026.

Release
28 SEP2026

RE-PoLar is out

Reproduction and extension of PoLar, with code and an interactive explorer.

Meeting
27 SEP2026

SWAMBIT at the first NI:KI project meeting

SWAMBIT introduced to the initiative, plus a team picture.

Poster
14 SEP2026

Poster at the NEAT workshop

The RE-PoLar poster, presented at the NEAT workshop.

in
Launch
1 DEC2025

SWAMBIT officially starts

SWAMBIT starts as one of the 11 NI:KI pilot projects.

NI:KI

Research focus

SWAMBIT builds AI that is functionally inspired by the brain: many small, sparsely active specialists instead of one dense monolith, coordinated through a shared workspace and enriched by a hippocampus-like memory. The goal is models that are more efficient and easier to extend.

  • Specialist Modulescortical macro-columnsSmall, interpretable transformers per modality and task; only a sparse few are active at a time.
  • Shared Global WorkspacethalamusA sparse write bottleneck where modules share one representational space.
  • Memory BenchhippocampusA hippocampus-inspired pointer memory, supplying factual knowledge on demand.
output
modalities
x1⋮xN
Shared Global Workspace (thalamus)
hierarchical, small Transformers (cortical macro-columns)
past experiences and knowledge (hippocampus)

memory
bench
sparse write-access
(bottleneck)
values (sparsely matched)
shared queries
umwelt
decision-action
system
(frontal
cortex)

Team

Publications

Programs-of-Layers in LLMs through the Lens of Cortical Areas RE-PoLar

Justus Westerhoff*, Stephan Olbrich*, Hatem Oraby, Walter Senn, Matthew Evan Larkum, Felix Gers

arXiv preprint arXiv:2609.31360, 2026 · *shared first authorship, random order

Cortico-subcortical multi-head self-attention as a substrate for cognitive performance

Vasanth Kumar Babu, Arno Granier, Timothée Proix, Titas Balvocius, Marmaduke Woodman, Ausra Saudargiene, Viktor Jirsa, Walter Senn

bioRxiv, 2026