The Lucy Syndrome — a reading path
Corrections don’t survive the next session, and more context doesn’t fix it. Centralizing memory, longer prompts, richer history — all of it sits inside the model’s reach, and the correction still gets overwritten on the next turn. What closes the loop is enforcement that runs outside the model. That is the through-line of everything below. Read it straight through, or jump in wherever you already are.
Start here
- 01 · The Lucy Syndrome: Why LLMs Forget Corrections — the diagnosis, in short.
The argument
- 02 · The Lucy Syndrome and AI — the five-part essay: definition, mechanism, and what a production system needs in order to actually learn.
- 03 · Questions and answers — scope, method, and the objections the work tends to provoke.
- 04 · Where this came from — the informal origin story behind the essay.
From idea to primitive
- 05 · From memory to scar: a four-layer progression — why memory stores are Layer 3, and hooks are the Layer 4 that closes the loop.
- 06 · Functional Scars — turning corrections into a primitive — the installable version:
pip install fscars.
Inside the lab
- 07 · A month of functional scars — 934 fires, one broken validation loop, and what the numbers forced me to build.
- 08 · The same scar, two agents — the same correction firing deterministically in both Claude Code and Codex.
- 09 · How do you know a correction held? — instrumenting an agent in production to tell whether a fix actually held.
What’s next
The series continues as the work does:
- Cross-agent enforcement, formally — why running outside the model makes the mechanism platform-agnostic, not a Claude Code trick. (~June)
- When the technical isn’t democratic — what happens when the statistically common answer is the wrong one in an expert domain. (~July)
- Ninety days of scars — the first real metrics report from the production system. (~July)
New here? Start at the top. Already building with hooks? Jump to Functional Scars or the Tools page.