New LLM Wiki Template Preserves Research Failures, Boosts Accuracy

Research projects routinely lose the most valuable parts of their process: dead ends, abandoned iterations, and retracted claims. These failures rarely survive into publications or shared code, forcing future collaborators to reattempt the same mistakes. Priscila Saboia Moreira and Christopher Sweet address this with llm-wiki-memory-template, an agent-aware instantiation of the llm-wiki pattern (Karpathy, 2026). The template inserts an LLM-maintained, interlinked wiki between raw sources and coding agents. Crucially, the wiki is append-only by design, preserving what didn't work alongside what did. This solves a negative-result loss problem that publications and code-sharing structurally cannot solve. The template supports heterogeneous collaboration along three axes: multi-human, multi-AI-agent, and multi-domain.

Three deployed case studies and one design report demonstrate the template's power. In a two-author project, a retroactive audit using the preserved failure path revealed that prior experiments' claimed 20-out-of-20 coverage was actually 14 and 12 evidence-based answers. After a fix guided by the preserved record, coverage reached 18 and 18. The failure path remained intact across the artifact. Other cases include a solo research lineage preserving abandoned iterations and an in-progress multi-agent deployment. The authors identify three cross-cutting sociotechnical properties beyond technical mechanisms: failure-path preservation, agent honesty, and appropriation. This substrate ensures that future researchers and AI agents can learn from every step, not just the final polished result.