BAML and Jev: building coding agent memory pipelines

Dex and Aaron walk through a memory pipeline developed by the Human Layer and Boundary teams. It extracts codebase facts, human corrections and recurring investigations from coding sessions, then proposes updates for CLAUDE.md or AGENTS.md. The aim is to preserve knowledge a coding assistant would otherwise have to rediscover in its next session.

The walkthrough shows how findings from different teammates and sessions become grouped recommendations with supporting references. A second filtering stage asks whether a finding adds information beyond what an agent can readily read in the code. In one example, Aaron says the report exposed an existing architecture document that agents were missing. A reference to that document could address the problem without duplicating its contents.

BAML defines structured inputs and outputs, while Jev handles classification and filtering before stronger models assess selected material. Aaron describes using Claude Code to explore the data first, then turning useful patterns into a more constrained workflow. Custom dashboards, cached requests and replay help him inspect prompts and compare results with an initial reference run. He also presents BAML query as tooling that can run locally; this does not establish that the model calls run offline.

Memory maintenance remains unresolved. The speakers discuss linking memories to code regions and using approvals, rejections and later corrections to identify stale advice and improve future recommendations.