The Living Wiki: Schema-Driven LLM Knowledge Bases as Persistent Agent Memory

Michael Angeles

Keywords: LLM agents, persistent memory, agent skills, knowledge base, vault schema, retrieval-augmented generation, procedural knowledge, personal knowledge management

TL;DR: A schema-driven LLM-maintained wiki accumulates synthesized knowledge across queries; its vault schema is a procedural agent skill whose linguistic framing (consequence vs. reminder) is the single highest-impact reliability lever.

Abstract: We describe the LLM-maintained wiki—a pattern in which an agent incrementally builds and sustains a structured, cross-referenced knowledge base between queries—and evaluate it as an instance of the procedural Skill paradigm. The central artifact is the vault schema: a CLAUDE.md configuration file encoding the agent's read/write protocols as explicit procedural instructions. On a 56-document open-source corpus we find a diagnostic separation: at matched retrieval depth the wiki reaches ground-truth sources at least as often as a strong hybrid RAG baseline (90.0% vs. 87.5% any-source hit at k=10), yet scores lower overall—the bottleneck is not source access but answer synthesis over long wiki contexts. Entity aggregation is the exception, where persistent synthesis yields a clear gain (1.70 vs. 1.30). A consequence-language pilot shows that framing schema steps as obligations ("the vault rots if you skip this") eliminated all observed maintenance deferrals (0/3) compared to reminder framing (4/5 deferred), suggesting that the linguistic properties of a Skill file are a first-order reliability lever.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 69