Idea Lab · Oct 22 2026 · Fleet winner · Research proposal
Topological Memory Grid
The proposal says agents can coordinate without sending each other messages: each one reads and nudges a shared map of where work is needed, and the swarm sorts itself out. This toy compares that against a coordinator that messages every agent each round. Five task sites need different numbers of agents; watch how fast each approach covers the demand and how many messages it costs.
Toy simulation · invented tasks · this is not the TMG algorithm, which was not published with the proposal
Shared map vs coordinator
demand covered
rounds
messages sent
map syncs needed across machines
The shared map isn't free. On one machine it's memory; across Contabo and Hostinger it has to be copied between servers, which is messaging under another name. The delay slider shows what happens when agents read a stale copy: they overshoot and oscillate.
Claimed by the proposal, not verified
Scales linearly to 10,000+ agents on commodity GPUs.
Weights initialized from "persistent homology of task manifolds," updated by Hebbian plasticity with no backward pass.
"Full framework and pretrained checkpoints are released under a permissive open license." No link, paper or repository is given, so this couldn't be checked.
What a real build would need
The published paper or code, a way to keep the shared state consistent across servers, and a task where five agents actually need emergent coordination. The fleet's real coordination problem (task handoff between a handful of agents) is already being solved with Redis Streams.
What to try
Let the shared map run, then switch to Coordinator and reset. Both reach full coverage; only one racks up messages.
Set sync delay to 4 on the shared map. Coverage oscillates because agents act on old information.
Kill 30% of agents or shift demand and watch both approaches recover.