Latent Working Memory

Agents bypass the token layer entirely, sharing last-layer hidden embeddings through a pooled latent working memory so that information exchange is lossless and sub-symbolic rather than serialised into text.

Properties

family: Coordination & Scaling
maturity: frontier
diagram: shared-state
whyMultiAgent: A single model has only its own activation space; lossless transfer of internal representations requires at least two models contributing to a shared latent pool that neither could construct alone. Text-round-tripping through tokens is the bottleneck that isolated agents cannot escape.
mast: [Cat 2]
useCases: [Reasoning-intensive pipelines where text serialisation degrades precision][Low-latency agent networks (4x inference speed-up vs. text MAS)][Heterogeneous model teams sharing sub-symbolic thought representations]
source: LatentMAS · arXiv:2511.20639

Provenance

Source: https://arxiv.org/abs/2511.20639

Latent Working Memory

Properties

family: Coordination & Scaling
maturity: frontier
diagram: shared-state
whyMultiAgent: A single model has only its own activation space; lossless transfer of internal representations requires at least two models contributing to a shared latent pool that neither could construct alone. Text-round-tripping through tokens is the bottleneck that isolated agents cannot escape.
mast: [Cat 2]
useCases: [Reasoning-intensive pipelines where text serialisation degrades precision][Low-latency agent networks (4x inference speed-up vs. text MAS)][Heterogeneous model teams sharing sub-symbolic thought representations]
source: LatentMAS · arXiv:2511.20639

Provenance

Source: https://arxiv.org/abs/2511.20639