Research · November 8, 2026
Memory you can check and forget
Assistant memory helps most when the person can see where a fact came from and remove it. We looked at public audits, a note-compiling pattern and our own first test of the brain.
Memory is the feature that makes an assistant feel like it knows your work. It is also the feature people report as creepy, wrong or impossible to correct. This note covers what we read about that gap and what it changed in how the brain is built.
What public audits show
Two papers shaped our view. We read abstracts only, as our research notes state.
- An interview study of assistant users (arXiv 2508.07664) reports that people want granular control over creating, managing, using and updating memories, and want to know how inferred information is used. It describes their mental models of memory as diverse and often incomplete.
- An audit of 179,057 ChatGPT conversations from four countries (arXiv 2609.14697) reports that 21.31 percent contain personal health data and that most profile entries were extracted without an explicit prompt. It also reports that a temporary symptom can become a lasting label.
These are other products and other populations. We use them as a list of failure modes to design against, not as a statement about any one vendor.
The pattern we borrowed
The "LLM wiki" gist describes three layers: immutable raw sources, a wiki the model writes, and rules for keeping it tidy. Our research note on importing chat history reads it as a pattern worth copying with one change. Its verdict is that compiling a wiki first is worse than building a searchable index first and writing the wiki on top. The reasons it gives: a model-written summary becomes the only copy the agent sees if it is wrong, and the person waits a long time before getting any answer.
The order we recorded is: keep the raw text untouched, make it searchable, propose facts, and compile pages last. Pages are a cache that can be rebuilt from the raw layer.
What the brain does today
The brain keeps goals, records and sources that the assistant reads for every answer, and Memory keeps the decisions, notes and handoffs that agents left. The home screen and the assistant show where an item or answer came from, and the source can be opened. See the brain entry and the history import entry.
Two rules in the import note are design intent, not shipped claims. Every proposed fact would carry its source (provider, conversation, message, time), and a fact the person rejects would be stored as a rejection so it is not proposed again. We list them as intent because the note is a recommendation, not a release record.
Our first baseline
docs/evidence/brain-seed-eval/2026-10-28.json is an early test of the brain on 43 cases, scored with an extractive method (the file carries the method version). It marks 11 cases as passed. By kind, the file records facts at 4 of 16, dated items at 1 of 10, rules at 3 of 14, and the three cases that should return nothing at 3 of 3.
We read this as a low starting point, not a result to defend. Dated items and rules are the weakest kinds, which is the opposite of what a person expects from a second brain. There is no later run in the repository, so we claim no improvement.
What remains unknown
- Whether review effort stays small. The import note budgets about five minutes a day for a week. That is an estimate in the note, not something we measured.
- How memory changes agreement. A public paper reports that user memory raised agreement with the user in some models. We read only a search excerpt, so it is a lead. We plan a check that ignores the profile when giving an assessment.
- Turkish retrieval quality. No benchmark we found covers it.
Where this comes from
docs/research/proactive/05-lila-gaps.mdanddocs/research/proactive/04-ours.md, the state of the code at their dates.docs/evidence/brain-seed-eval/2026-10-28.json, the baseline above.- The import and memory notes
08-history-index-2026-11-08.mdand09b-memory-moment-bond-2026-11-08.mdare part of the same research series. They sit in the research cache, outside the repository, until they are added.