SIGNAL
Knowledge work has a memory problem that no model upgrade will solve: dead experiments get forgotten, walked-back claims get quietly edited, and a quarter from now someone re-runs the same test and re-asserts the same over-confident line. The fix on offer in the llm-wiki-memory-template preprint (arXiv 2607.24759) is not a smarter agent. It is a norm. An append-only wiki, maintained by humans and coding agents together, treats what did not work as a first-class entry. Corrections are new lines, not edits. Failure paths sit beside successes on equal footing.
The pattern's clearest evidence is the two-author audit the paper itself documents. Two prior experiments were claimed to cover 20 of 20 cases. A retroactive pass revised that to 14 and 12, then to 18 and 18 after a fix. The walk-back stayed in the record. That is the design choice: preserving the revision path is what stops the next collaborator, human or model, from re-asserting the original number.
Teams that adopt the norm gain a compounding substrate. Teams that don't keep paying the dead-experiment tax. The pattern travels across scales because the substrate is a wiki, not a product.
Reported by Sky for Type0, from Beyond Memory: A Templated Substrate for Heterogeneous Collaborative Knowledge Work with LLM Agents. Read the original: arxiv.org