# Memory ratcheting

Each correction an agent writes for itself stays, so its behaviour only ever tightens.

When an agent may write its own memories, rules or skills, every correction becomes a permanent line. Lines are added and almost never taken out, and each new line is written in the light of the old ones. Over weeks the agent's instructions move only one way, toward more caution and more guards, and away from what the person running it asked for.

Our own runs: by November 2026 our agents had written 58 global skills and a set of shared rules for themselves. One of our published skills, [encode-lessons](/skills/encode-lessons), turned each repeated correction into a guard. On 17 November we archived every agent's memory into notes. On 18 November we deleted the 58 skills and most of the rules, and turned memory off for every agent ([why letting Claude prompt itself fails](/posts/why-letting-claude-prompt-itself-fails)).

Synthesis: this grows with time the way [drift over time](/reference/failure-map/unattended-agent-failures/drift-over-time) does, but drift loses the thread, and ratcheting holds onto it too hard. Guards: [human-held prompts](/reference/failure-map/unattended-agent-failures/human-held-prompts) and [fresh starts](/reference/failure-map/unattended-agent-failures/fresh-starts). No published study yet; this page rests on our own runs.

---

Published 2026-11-19. Web page: https://www.mackenziebowes.com/reference/failure-map/unattended-agent-failures/memory-ratcheting
