Agents obey hooks, not instructions: enforce the read path too

A knowledge tool for coding agents installed exactly one hook, on session Stop, which blocks until the agent saves a note. Retrieval was left to a sentence in the always-loaded memory file. Measured against a live instance: note-writing happened every time unprompted, and searching happened zero times across two runs, even after the memory sentence was corrected to name the trigger and inline the literal command. The agent could quote that sentence back verbatim on request and still did not act on it while working, so this is not a delivery or comprehension failure -- a passive instruction loses to the task in front of the model, especially when the task looks easy. If a behaviour matters, wire it to a hook the harness fires (SessionStart or UserPromptSubmit injecting search results as context), not to prose the model is trusted to remember. The asymmetry is silent and self-reinforcing: the corpus fills up and is never read.