Re-explaining your life to an AI every session is evidence that the system is not paying attention. Personalised AI should learn continuously, act proactively, and never leave your device.
What we ruled out
Frontier 3 is Anthril’s longest bet. Most consumer AI assistants route everything a user types to a hyperscaler context — stateless, amnesiac, and expensive. We are building the alternative.
A cultivant is not a user who prompts a service. It is a person who grows an AI through ongoing interaction. The AI learns their routines, preferences, relationships, and goals. The cultivant shapes the AI; the AI shapes the cultivant’s day. The relationship compounds over time in ways that stateless AI cannot approximate.
Re-explaining your context to an AI every session is evidence that the system is not paying attention. The problem is not insufficient prompting — it is that stateless systems have no continuity of experience. They cannot notice you have not planned dinner for the week, check what ingredients are available, order what is missing, and tell you what to pick up. They cannot flag that Thursday’s appointment clashes with a school pick-up you mentioned in October.
Like Frontier 2, Frontier 3 products use existing frontier LLMs as the reasoning engine. What we layer on top are Aurora-derived memory and learning systems: multi-tier personal knowledge graphs, event schemas for capturing life events, episodic memory that persists across sessions, and consolidation cycles that compress patterns over time — without sending personal context to a cloud provider.
Memex is the working hypothesis — a local desktop knowledge base that runs on-device, learns through interaction, and grows more useful with every week. v0.2.0 is shipping now.
Open questions
Q01
H3.1 tests this. The hypothesis is that continuity of experience is a primary driver of perceived AI value — and that a cultivant whose AI knows their routines, preferences, and constraints will rate it significantly more helpful than a stateless assistant on the same tasks. Memex is the working bet.
Investigated by: memex
Q02
H3.2. The hypothesis is that privacy-by-architecture unlocks a class of personal information that cloud systems cannot access precisely because they are not trusted with it. If confirmed, local-first is not a constraint — it is a competitive advantage.
Investigated by: memex
Q03
H3.3 tests proactive agency with read/write access to daily life systems. The key variable is calibration — an AI that acts on patterns but cannot be tuned to individual tolerance thresholds will generate intrusive over-automation. The hypothesis is that calibrated proactivity is both achievable and measurably better than reactive prompting.
Investigated by: memex
Currently open
Memex v0.2.0 is shipping with 13 new features. The H3.1 longitudinal study design is in progress. Aurora's H3 and H4 hypotheses will be validated through Memex usage in production.
Related publications
