Beyond Memory: A Templated Substrate for Heterogeneous Collaborative Knowledge Work with LLM Agents

Summary

The article proposes a template called 'llm-wiki-memory-template'. This template functions as a wiki managed by a large language model (LLM) that only allows appending. Positioned between raw sources and AI agents, this structure preserves structurally handled failure paths, dead ends, and retracted claims from publications and code repositories. The template is positioned as infrastructure for knowledge work based on multi-human, multi-agent, and multi-domain collaboration.

Why it matters

The article targets a real and sufficiently unaddressed gap. Negative results and abandoned attempts disappear from research records. This leads to repeated failures across teams and over time. The wiki's append-only nature and agent awareness create an accumulative knowledge layer that RAG over raw sources or transient agent session memory cannot provide. A case study where a retrospective audit corrected exaggerated experimental scope claims (dropping from 20/20 to 14/12, then to 18/18 after a correction) serves as concrete, albeit small-scale, evidence of the value.

Highlights

• The append-only wiki tradition preserves failures alongside successes. This approach directly addresses the 'loss of negative results' problem that publication norms fail to resolve.
• The template is framed as a sociotechnical product. 'Agent honesty' and 'adaptation' are named as first-class features, not just technical mechanisms.
• A two-author case study shows that a retrospective audit caught and corrected an exaggerated claim of 20/20 full coverage. The complete failure path is preserved in this product.

Critical take

The experimental foundation appears weak. The 15-page text contains three case studies and one design report. The multi-agent aspect is treated as a design sketch rather than a deployment example. The underlying 'llm-wiki pattern' is attributed to Karpathy (2026) and an informal source ('tonbi, 2026'). This raises questions about how much of the contribution is genuinely new versus a tidy repackaging of existing community practices.