Wiki Foundation Model retrieves agent memory from linked markdown wikis

The Wiki Foundation Model (WFM) is for agents whose long-term memory is a folder of linked markdown files. It turns an LLM wiki into a graph and retrieves with message passing conditioned on the query, so the dense text on each page and the links between pages shape the result together.

The team also built a GPU-to-GPU training protocol that it says trains 10.5 times faster, and it reports strong results on five benchmarks covering agent memory and multi-hop reasoning. The work is described in the WFM paper, and academy.dair.ai hosts a chat page for the paper.