---
title: "Chapter 15: The No-Weights Complement"
description: "Understand Reflexion notes and skill accretion as zero-GPU learning, check the idea, then wire a notes file and one skill into the harness."
type: lesson
order: 15
chapter: "05-the-flywheel"
---

# Chapter 15: The No-Weights Complement

## Objective

Add learning mechanisms that run **in parallel** with weight updates and cost **zero GPU**: Reflexion-style notes and skill accretion.

---

## Concept

### Reflexion-style notes

After each run the agent writes a lesson to a persistent file; next run reads it. Instant learning, no training.

### Skill accretion

When the agent solves something reusable, it writes a script or `SKILL.md` and keeps it. Experience becomes tooling.

### Why this belongs in a fine-tuning course

Fine-tune → competent at **the loop**. Notes/skills → competent at **your repo, this week**. Different timescales, no conflict.

This is also the resolution of “don’t bake codebase facts into weights”: notes and skills are the deliberate context side of that split.

---

## Concept check

**Q1.** What kinds of knowledge should go into notes/skills rather than LoRA weights?

**Q2.** How do notes and fine-tuning complement rather than replace each other?

<details>
<summary>Concept answer key — attempt first</summary>

## Answer key (concept)

### Q1

**Model answer:** Repo-specific, fast-changing facts, local conventions, one-off caveats — things that go stale in weights.

**Pass criteria:** volatile / repo-specific knowledge

### Q2

**Model answer:** Weights hold durable loop/protocol habits; notes hold current project context — both run together.

**Pass criteria:** loop vs context timescales

</details>

---

## Gate: Notes file in the harness

- After each run (or each failure), append a short lesson to `memory/notes.md` (or similar)  
- Prepend or inject notes into the next system/user context  

**Pass:** two sequential runs; second run’s prompt includes a note written after the first.

---

## Gate: One skill artifact

When a reusable fix appears, write `skills/example-SKILL.md` or a small script and teach the agent (or yourself) to load it.

**Pass:** skill file exists; harness or docs reference how it is loaded.

---

## Check your understanding

**Q3.** Why does this chapter cost “zero GPU” but still improve agent quality?

**Q4.** If you only fine-tune and never take notes, what class of failures keeps recurring?

<details>
<summary>Answer key — attempt every question first</summary>

## Answer key

### Q3

**Model answer:** It changes context/tools the agent sees, not weights — still improves next-run behavior.

**Pass criteria:** context/tooling not training compute

### Q4

**Model answer:** Project-specific gotchas and repo facts the weights should not memorize.

**Pass criteria:** recurring local/context failures

</details>

---

[← Chapter 14](/fine-tune-local-agent/05-the-flywheel/14-learning-from-failures/) · [Next: Chapter 16 →](/fine-tune-local-agent/05-the-flywheel/16-scale-up-and-ship/)