Definition
Plain language
The idea that a note you can read and second-guess is safer than a tool you just run and trust.
As stated in the literature
In Metis's text-versus-code memory analysis, the observation that text experience is consumed as adjustable advice a model can filter against reality, while code is invoked as a trusted black box — so a flawed code memory propagates its defect to every caller and suppresses the agent's own error-recovery behavior.
Why it matters: It explains why storing an agent's experience as adjustable notes can be safer than storing it as trusted code that silently spreads its defects.
For example, an agent can weigh a written tip against what it actually sees and ignore it if it's wrong, but it will run a saved snippet of code as-is and inherit any hidden flaw.
Heard on the show
“And the authors give it a name — the injection asymmetry — and honestly they hand you the perfect analogy for it.”Episode 168 — When Turning Experience Into Code Makes Your AI Agent Dumber