Theo Browne argues that source code should remain the ground truthSource code as a source of truth means treating the current implementation and its executable constraints as the authoritative record of software behavior. for coding agents. Separate memory systemsAI agent memory is stored information that an agent can retrieve and use across steps, sessions, or changing contexts. create another store that can drift from the implementationAI agent memory drift occurs when stored agent knowledge no longer matches the current code, environment, policy, or task state., while modern agents can usually inspect a repository directlyRepository context is the relevant source, configuration, history, documentation, dependencies, and worktree state needed to understand a software task. with ordinary shell and search tools.
Theo Browne audits the memories accumulatedAn AI agent memory audit reviews stored entries for relevance, provenance, accuracy, duplication, usage, scope, and expiry. by Claude Code across his own projects and finds temporary plans, finished fixes, benchmark details and duplicated project rules. Many entries were written but never read, and several could steer future work using information that was already obsolete.
Theo Browne recommends encoding durable direction in a concise AGENTS.md or similar project guide, then preventing repeat failures through clearer architecture, lint rules, tests and CI. Skills and human review remain useful fallbacks, but automatic memory should not replace a codebase that makes its values and constraints explicit.
Watch on YouTube
