To represent you, an AI agent has to understand you. Here, an agent means any AI assistant that writes, plans or acts for you. It needs facts about you. It also needs to know how you interpret those facts and how you tend to make decisions.
You can give it that information in two ways. You can write a profile of how you think, or you can build one from your existing writing. This page covers both, with a worked example.
What memory and custom instructions already do
ChatGPT and Claude both let you save memories and give the assistant standing instructions. These work well for facts and preferences, such as your job, your tone, the tools you use and the format you like.
They hold what you have told the assistant about yourself. They do not usually describe how you weigh a tradeoff, what you consider settled, or how your actions change in different situations. Most people have never written that down, and it is hard to write from memory.
Option 1: write it down yourself
Open a plain text file and answer questions like these in your own words.
- What do you check first when you make a decision?
- When two good options conflict, which one usually wins, and when does that change?
- What kind of evidence convinces you, and what kind does not?
- How do you want disagreement raised with you?
Save it as a personal context file. Paste it into your custom instructions or a project's instructions. Coding agents read files such as CLAUDE.md or AGENTS.md, so you can put it there too. You need no software for this. The limit is that you can only write what you already know about yourself.
Option 2: derive it from your writing
Your chats, essays, notes and transcripts show how you reason. Base Layer is an open-source pipeline that reads your writing and creates a specification for AI agents. The specification contains plain-text claims about how you reason, communicate and decide. An agent reads it before acting for you.
Base Layer uses the cited build as its method. Every claim cites the facts it rests on, and every fact quotes your own words. You can trace a claim back to what you said and correct it. baselayer run is the older, quicker path, with fewer checks.
Aarik Gulaya, who built Base Layer, described the idea this way: "if an agent is meant to represent you, then it needs to have an understanding of you, not just facts on you, but how you interpret those facts".
A worked example
This claim comes from a specification built from Zitkala-Sa's American Indian Stories (1921, public domain). P15 is the claim's id, and P marks the PREDICTIONS layer.
P15 TOLERANT TO PEOPLE HARSH TO SYSTEMS. Their severity is aimed at institutions and at settler power in the aggregate, not at individuals in front of them: in face-to-face disagreement, including religious disagreement, they extend respect, pity for suffering, and forgiveness over punishment. Predict the polemic in writing and the courtesy in the room.
Active when: They disagree with, or witness the hardship of, a particular person present
Evidence: [F-86fc472d] [F-cabf4c6c] [F-124a3cd5] ... (17 facts)
A memory system would store facts from her writing. This claim tells an agent how to interpret and apply them. Her essays are severe about institutions, and the claim says she is courteous to the person in front of her. The Active when line says when the claim applies, so an agent does not apply it everywhere. The Evidence line lists the ids of the 17 facts used to write it.
This published example used an earlier version of the cited build, before facts quoted their source text. Its evidence therefore shows fact ids only. In the current cited build, each fact also quotes the words it came from. The full specification is in the repository.
How to try it
Base Layer needs Python 3.10 or later and an Anthropic API key. The current version is 0.5.0, and you install it from GitHub.
The cited build takes eight commands, listed in the README. It writes files that you add to an agent's instructions yourself. For about 50,000 words of your writing, the README projects $5 to $7.
To try Base Layer quickly, use baselayer run. Its claims can cite facts but do not have to. The README projects $1 to $2 for the same amount of writing.
pip install git+https://github.com/agulaya24/BaseLayer.git
baselayer run chatgpt-export.zip
claude mcp add --scope user --transport stdio base-layer -- baselayer-mcp
The middle line builds the specification from a ChatGPT export. The last line connects the result to Claude Code through MCP (Model Context Protocol). Claude Code is Anthropic's coding agent.
Base Layer sends your text to the Anthropic API. Extraction is the step that pulls facts out of your writing. It can run on a local model through Ollama instead. Base Layer stores the database, facts and specification only on your machine.
What to keep in mind
A model writes the claims, and a model can be wrong. Only you can confirm that a claim describes you. Read the specification and edit anything that is wrong before an agent relies on it.
The research behind Base Layer tested specifications on 14 historical figures who wrote autobiographies. Panels of AI models served as judges. Specifications helped most where the model knew the person least. The paper is on arXiv.
A later test, separate from the paper, used recorded responses to the same decisions under nine specifications. A simple classifier picked the right specification 51.6% of the time from the reasoning and 13.4% of the time from the decision alone, against 11.1% by chance. The README sums it up this way.
"Specifications changed how decisions were argued in every situation tested. They changed the decision itself in some situations."
Source code: github.com/agulaya24/BaseLayer.