When an AI misreads you, it may not have forgotten anything. It may never have known how you think.
An assistant with memory can hold many facts about you. It can know your job, your projects and how you like answers formatted. It can still give advice you would never take or write in a way you would never write. In that case, it had the facts. It was missing how you interpret them.
Knowing facts about you is only part of it
Two people can know the same facts and make different decisions. Aarik Gulaya, who built Base Layer, puts it this way: "memory is really about how the facts are used. Why leave that to a pure inference machine, you must tell it".
Memory features store the facts and preferences you share or that the assistant picks up. Few people tell an assistant what they count as good evidence, which concerns matter most to them, or how they behave in different settings. When the AI has to fill that gap, it has little to go on beyond what people generally tend to do.
Three common ways it goes wrong
It applies a preference everywhere. You may be blunt in writing and gentle in person. An AI that learned "blunt" can end up being blunt in every situation.
It smooths over your contradictions. Most people hold views that pull against each other. An AI may pick one side or blend both into something you never said.
It knows public figures better than it knows you. A model has read a great deal about famous people and very little about you. In the research behind Base Layer, a written specification helped most for the people the model knew least. For the 9 of 14 historical figures the model knew least, every one scored higher with a specification and facts than with no context. The average increase was +0.89 on a 1 to 5 scale.
What can help
Give the AI a description of how you reason, in words you can read and correct. You can write one yourself in a personal context file or in your custom instructions. You can also create one from your own writing, which is what Base Layer does.
Base Layer is an open-source pipeline. It reads your chats, essays, notes or transcripts. Then it writes a specification, a set of plain-text claims about how you reason, communicate and decide. An AI agent is an assistant that writes, plans or acts for you. It reads the specification before acting for you.
Base Layer's cited build lets each claim state the situation it applies to. That helps with the first problem in the list above. When your own words disagree, the build marks the claim as CONTESTED and keeps both sides. It tells the agent to leave the conflict unresolved. That helps with the second problem. baselayer run is the older, quicker path, with fewer of these checks.
An example of a claim with a condition
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)
An AI that only knew her published essays might expect her to be harsh with everyone. The claim says her severity in writing is aimed at institutions. With a particular person in front of her, she tends toward respect and forgiveness. The Active when line says when that applies. The Evidence line lists the ids of the 17 facts used to write the claim. In the same specification, claim P2 is marked CONTESTED.
This published example used an earlier version of the cited build, before facts included quotes from 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 example is in the repository.
Limits
A model writes these claims, and it can misread you too. Only you can confirm that a claim describes you. Read the specification and correct it before an agent relies on it. Base Layer runs on Anthropic's models. The step that extracts facts from your writing can run on a local model through Ollama instead.
The research is in the paper on arXiv. It tested specifications on 14 historical figures, with panels of AI models as judges. The code, the example and setup steps are on GitHub.