WFM trains on linked markdown wikis for agent memory, with a 10.5x faster GPU-to-GPU protocol

omarsar0 · x · 2026-09-23

A new paper introduces the Wiki Foundation Model (WFM): an agent's long-term memory stored as linked markdown pages is turned into a graph, and retrieval uses message passing conditioned on the query, so both page text and link structure shape results. The team also built a GPU-to-GPU training protocol that trains 10.5x faster, reporting strong results across five agent memory and multi-hop reasoning benchmarks. Directly relevant to anyone keeping agent memory as folders of linked markdown files.

Original post →

More from coding & agent

coding & agent channel →