A new arXiv paper argues that research teams lose too much useful knowledge because publications and code usually omit false starts, abandoned claims, and reasoning paths. The authors propose an LLM-maintained wiki template as a shared memory layer for collaborative knowledge work.
The pattern places an interlinked wiki between raw sources and the agent. It is designed to be append-only by convention, so failed attempts and walked-back conclusions remain visible instead of disappearing from the record.
The authors present case studies covering solo research, two-author collaboration, multi-agent work, and an educational variant. In one audit, the wiki helped revise earlier claimed coverage from 20 of 20 answers down to evidence-based counts of 14 and 12, then up to 18 and 18 after fixes while preserving the failure path.
The work is not just about memory length. It treats persistent, inspectable records as a way to make agent work more honest, reusable, and easier for future collaborators to challenge.