korrents

korrents · piece

Your AGENTS.md is a Neural Net

Kun Chen · 23 Aug 2026

Read the piece on blog.kunchenguid.com ↗

3 korrents from this piece ↓

Kun Chen did not write this page.

Every claim below was made in this piece, quoted word for word and numbered in the order the piece makes them, so you can read it there rather than take our word for it. The sentence above each quote is our reading of the claim, not their wording. Each quote was checked against a stored copy of the page at build time; where the two differ, the quote is the fact.

  1. Our reading

    An agent memory file that keeps growing dilutes how well the agent follows instructions, burying the rules that matter under ones that don't.

    Their wordsWorse, instruction following gets diluted as the file grows, so the rules that matter are buried under rules that don't.

  2. Our reading

    A project-level agent memory file works like a neural net, steering the agent much as a fine-tuned LoRA steers a model.

    Their wordsThe project-level file is actually a neural net. It gets loaded into the system prompt of agent sessions that run on the project, and it steers the agent's behavior not too differently from a fine-tuned LoRA that applies on top of the model you use.

  3. Our reading

    Changes to an agent's project instructions should be driven by the agents' session logs, not by what the author remembers being annoyed by.

    Their wordsThe loss signal should be from the agents' session logs, not your recollection of being annoyed last Tuesday.

Keep reading