LLM Knowledge Bases: a practical guide — Ben Holmes, Warp
A useful agent knowledge base starts with abundant raw Markdown, then adds controlled tags, backlinks, and generated indexes. Scheduled agents can maintain it, but sync and taxonomy need explicit rules.
Holmes starts with **raw Markdown notes**, often captured by voice, then uses an enrichment skill to add sources, timestamps, tags, and backlinks. A fixed tag reference keeps agents from creating a new taxonomy on every pass.
Give coding agents a staged pipeline: capture first, enrich untouched notes, generate a navigable wiki, then build visualizations only when the corpus is substantial. A **scheduled cloud sandbox** can sync the folder, run skills, and return the updates.
Holmes starts with **raw Markdown notes**, often captured by voice, then uses an enrichment skill to add sources, timestamps, tags, and backlinks. A fixed tag reference keeps agents from creating a new taxonomy on every pass. Give coding agents a staged pipeline: capture first, enrich untouched notes, generate a navigable wiki, then build visualizations only when the corpus is substantial. A **scheduled cloud sandbox** can sync the folder, run skills, and return the updates. Generated links and wikis inherit errors from the notes and enrichment pass. Local automations also require the laptop to be available, while cloud execution introduces note-sync and data-handling decisions.
This provides a lightweight implementation path for durable context: preserve raw notes, enrich them through a controlled taxonomy, and derive navigation only after capture. It makes shared context practical without requiring a new memory architecture, while confirming that generated structure is not authority: provenance, taxonomy discipline, scheduling, synchronization, and data handling remain operational requirements.