Author(s): Codebook Fusion

Originally published on Towards AI.

I used to think a bigger context window would fix my agent’s memory problem. It didn’t. It just made the mistake more expensive.

Every task started with the same ritual. The agent opened a handful of files, worked out which module owned what, and rebuilt a mental map that it had already built an hour earlier. Then the session ended and the map disappeared.

After the intro, the author explains how building an always-updating codebase knowledge graph avoids repeatedly re-learning structure, using Google’s Open Knowledge Format (OKF) where knowledge lives as a folder of linked Markdown files with YAML frontmatter. They describe why OKF is a good fit for code modules/services/imports, and share benchmark-related reasoning about token savings, including why headline multipliers can be misleading. The core design is laid out as three rules: extract deterministic facts from the code parser (AST) rather than hallucinating with a model, have the model write prose meaning only when needed via a gating scheme, and ensure everything is validated by a linter before committing. They then walk through the pipeline (diff → map → hash compare → AST extraction → conditional drafting → lint → commit), show how source_hash and surface_hash control what work is redone, and discuss a linter that rejects broken links, missing metadata, and nonexistent resources. Finally, they compare their approach to existing tools, outline honest downsides (small repos, CI noise, prose drift, refactor churn, and questionable single-project benchmark claims), and give guidance on when building one is worth it—emphasizing that the discipline of parser facts + model meaning + linting beats relying on format alone.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI