Abstract
Research projects, educational efforts, and adjacent knowledge work accumulate findings, decisions, and reasoning that future collaborators rarely recover. The most valuable parts of that work—including dead ends and retracted claims—are routinely excluded from publications and shared code. Consequently, future researchers often repeat the same failures because no record survives. While LLM coding agents are now common participants in such work, they lack persistent memory across sessions, and retrieval-augmented generation over raw sources does not effectively compound knowledge over time.
The "llm-wiki" pattern (Karpathy, 2026; tonbi, 2026) addresses this limitation by inserting an LLM-maintained, interlinked wiki between raw sources and the agent. In this paper, we present llm-wiki-memory-template, a reusable, agent-aware instantiation of this pattern. We argue that it serves as a substrate for heterogeneous collaborative knowledge work along three dimensions: multi-human, multi-AI-agent, and multi-domain. Each dimension is supported by a distinct architectural element of the template (§4).
The wiki operates on an append-only convention, preserving what did not work alongside what did—thereby addressing the problem of negative-result loss that publications and code-sharing structurally cannot solve. We present three deployed case studies and one design report covering these dimensions individually:
- A solo research lineage that preserves abandoned iterations.
- A two-author project whose retroactive audit revised two prior experiments' claimed 20-of-20 coverage down to 14 and 12 evidence-based answers, then to 18 and 18 after a fix—with the failure path preserved across the artifact.
- An in-progress multi-agent deployment reported as a design.
- A cross-domain educational variant.
We identify failure-path preservation, agent honesty, and appropriation as cross-cutting sociotechnical properties of the artifact—not merely of its technical mechanisms.
Keywords
LLM agents, collaborative knowledge work, failure-path preservation, wiki-based memory, heterogeneous collaboration
via ArXiv AI