Markdown is 40× more token-efficient with SQLite

Suggests structuring agent context in SQLite can cut tokens per correct answer by about 40x versus Markdown files. Typed frontmatter is a workable middle ground.

• In the founder's own 120-trial test over 11,000 records, agents used 40x fewer tokens per correct answer querying SQLite than reading the same data as Markdown files; the dataset is private, so the result is not independently verifiable.
• Agents still did reasonably well without SQL when Markdown files had strongly typed frontmatter: in that arm of the trial they wrote Python scripts to run SQL-like queries over the files.
• The author found cached input was over two-thirds of the token bill on long agent sessions, so letting sessions run long because cached tokens are cheap still adds up.

• token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
• context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

• videoYour company brain will leak secrets: how we stopped it for big banks — Tanmai Gopal, PromptQLAI Engineer
• postAgents rebuilt SQLite in Rust from the manual aloneCursor
• articlesparqlite: SPARQL 1.1 inside SQLite, written in Rustsimonwillison.net

Checking sign-in…

Loading comments…