Researchers propose llm-wiki-memory-template to preserve failure paths in collaborative AI and human workflows

Researchers argue that current publishing and code-sharing practices systematically exclude dead ends and walked-back claims, forcing future collaborators to repeat prior failures. LLM coding agents, while increasingly common, lack persistent memory across sessions, and retrieval-augmented generation over raw sources does not compound learning over time.

The proposed llm-wiki-memory-template inserts an LLM-maintained, interlinked wiki between raw sources and agents, enabling an append-only record that preserves what did not work alongside what did. The authors position this as a substrate for heterogeneous collaborative knowledge work spanning three axes: multi-human, multi-AI-agent, and multi-domain collaboration.

Each axis is supported by a distinct architectural element of the template, as detailed in the paper’s fourth section. The append-only convention is intended to address a "negative-result loss problem" that publications and code-sharing structurally cannot solve.

Three deployed case studies and one design report illustrate the template’s use. In one two-author project, a retroactive audit revised prior experiments’ claimed 20-of-20 coverage down to 14 and 12 evidence-based answers, then to 18 and 18 after a fix; the failure path was preserved across the artifact. Other case studies include a solo research lineage preserving abandoned iterations and an in-progress multi-agent deployment reported as a design.

The authors identify failure-path preservation, agent honesty, and appropriation as cross-cutting sociotechnical properties of the artifact, emphasizing that these properties arise not only from technical mechanisms but also from the artifact’s social embedding in collaborative workflows.