One tutor, many minds
Most AI helpers are reactive: you ask a question, they answer, and then they forget. A real tutor works very differently. They notice what you keep getting wrong, decide what to work on next, change tack when you are stuck, set the right practice, and check that you have genuinely mastered it. That loop, sustained over weeks, is what moves a learner forward, and a single-shot chatbot cannot hold it.
Project Cognita is ILM AI's research into closing that gap: multi-agent cognitive orchestration, a coordinated team of specialist agents that together behave like one attentive tutor.
From a goal to a plan, teach, assess, adapt loop
Give Cognita a goal, for example "get me ready for the algebra paper", and it runs the whole learning loop on its own: plan, teach, assess, adapt. It does not wait to be prompted one step at a time; it chooses the next best teaching action from what it currently understands about the learner.
At the centre sits Raphael, the orchestrating agent. Raphael does not try to do everything itself. It directs a team of specialists, each owning one job a human tutor does:
- Planner builds and continually revises the study path towards the goal.
- Explainer teaches the current idea, choosing the right example, analogy or diagram.
- Assessor marks work, probes for misconceptions, and checks for real mastery rather than a lucky answer.
- Memory carries what has been learned about that specific student into every future session, so progress compounds.
Because the work is split across specialists and coordinated by Raphael, each part can be as sharp as a focused tool, while the learner only ever sees one calm, coherent conversation.
Key capabilities
Goal-conditioned planning
A target grade or a deadline becomes a personalised, self-revising study path. In research terms, we are learning a policy over teaching actions, written as π(a | s, goal): given the learner's state and their goal, what should the tutor do next.
A coordinated agent team
Planner, explainer, assessor and memory agents work behind a single conversation, conducted by Raphael. Improving or adding one specialist lifts the whole tutor.
Purposeful tool use
Cognita calls on the wider platform when the moment calls for it: practice sets, interactive simulations, marking, and diagram generation (including our Lumina sketch engine). Tools are used deliberately, not sprayed at every turn.
A Socratic policy
Cognita guides with questions first and reveals answers last. The goal is understanding that the learner builds for themselves, not answers handed over.
Longitudinal memory
Mastery is tracked across sessions, so pace and strategy adapt over time rather than resetting with every conversation.
What we are researching
Cognita is an active research programme, not a finished feature. The questions we care about most:
- Cognitive architectures: how best to divide tutoring into agents, and how they should communicate, so the whole is genuinely more capable than one large model prompting itself.
- The next best teaching action: choosing what to teach, test or revisit from a live, uncertain model of the learner's understanding.
- Safety and honesty: keeping autonomous agents aligned, so they never fabricate, never over-help, and always leave the thinking with the learner. Human oversight stays central.
Where Cognita fits
Cognita is the orchestration brain behind ilmino. It connects the platform's AI tutor, revision planner, practice and analytics into one companion that can run a whole learning journey rather than answer one question at a time.
For an institution, that matters in two ways. It offers every student the kind of sustained, adaptive attention that is impossible to staff by hand, and it does so with teachers and leaders able to see, guide and override what the agents do. The research feeds the same platform your staff already control; it does not replace their judgement.
Cognita is early, and we are deliberately careful with autonomous tutoring. We would rather ship a tutor that is honest and occasionally cautious than one that is confident and wrong.
To see how Raphael and the wider platform work in practice, book a school demo. You can read more about our research at ilmai.co.uk.
