MindForm — Giving AI Agents a Persistent Identity

Inspiration

AI has become remarkably intelligent, but most AI agents still lack something fundamental: a persistent identity.

Today's language models can hold conversations, solve complex problems, and adapt to instructions. Yet their personalities often exist only in prompts. Change the context, restart a conversation, or introduce new experiences, and the agent may behave like an entirely different character.

We started MindForm with a simple question:

What if AI agents could develop a personality through experience, preserve their identity over time, and evolve without becoming inconsistent?

We believe the next generation of AI agents needs more than intelligence and memory. It needs an identity layer that determines how experiences shape future behaviour.

What MindForm Does

MindForm is a modular personality architecture that gives AI agents persistent, evolving identities.

Rather than defining personality through a fixed system prompt, MindForm represents an agent through interconnected psychological and cognitive components.

Its architecture includes:

Personality Traits: Continuous Big Five (OCEAN) traits that evolve gradually through experience.
Long-Term Memory: Persistent representations of experiences and relationships that influence future interactions.
Motivation and Drives: Internal motivational states that influence decisions and behaviour.
Self-Concept: An internal representation of how an agent understands its own identity.
Cognitive Patterns: Mechanisms that interpret incoming experiences through the agent's existing personality and beliefs.
Feedback Learning: A process for incorporating the consequences of the agent's own actions into its evolving state.

Together, these components help an AI agent behave as a recognisable individual rather than an interchangeable model responding only to the latest prompt.

How We Built It

MindForm separates language-model inference from persistent personality state.

Our research began with StableMind, a two-timescale architecture designed to address long-term personality drift.

A fast layer handles immediate conversational context, while a slower layer maintains established personality characteristics. Feedback mechanisms selectively update the persistent state instead of allowing each new interaction to overwrite it.

We subsequently developed a more modular architecture that represents personality, motivation, memory, self-concept, and behavioural feedback as interacting systems.

The implementation combines Python-based agent logic, language-model integrations, structured state management, and feedback mechanisms.

We have developed an interactive demonstration that allows people to experience personality-driven AI agents.

Challenges We Faced

1. Personality drift

A major challenge was preventing recent conversations from overriding an agent's established identity.

We addressed this through separate fast and slow adaptation mechanisms, allowing agents to respond flexibly without immediately rewriting their underlying personality.

2. Stability versus evolution

An agent with an unchanging personality feels artificial, but one that changes constantly lacks continuity.

MindForm explores this trade-off through gradual state updates, bounded personality representations, and feedback mechanisms.

3. Coordinating cognitive components

Personality is not a single number or prompt. Memory, motivation, beliefs, values, and self-concept interact.

Designing these as separate but interconnected components required careful attention to how information moves through the system and changes over time.

What We Learned

Building MindForm has reinforced an important distinction: remembering information is not the same as maintaining an identity.

An agent may remember a user perfectly while still behaving inconsistently.

We also learned that personality cannot be evaluated through isolated responses alone. Meaningful evaluation must examine behaviour across multiple interactions, changing circumstances, and extended periods.

These insights have shaped our ongoing research into personality preservation, expression, behavioural consistency, and evolution.

Hackathon Development

[Describe the specific significant improvements made to MindForm between 26 August and 30 October 2026.]

[Identify the NVIDIA open-source model integrated into the working application.]

[Explain how the application uses Nebius Token Factory or Nebius AI Cloud at runtime.]

[Describe what the integration enables and what you demonstrated or measured.]

What's Next

Our goal is to make persistent personality a reusable infrastructure layer for intelligent systems.

We are exploring applications across:

AI companions with consistent, developing identities.
Games and virtual worlds featuring characters shaped by their experiences.
Robotics where personality can influence interactions and adaptive behaviour.
Personal AI assistants that retain a recognisable identity over long-term use.

Our next research priority is systematic evaluation: measuring how personality persistence, memory, and adaptive behaviour influence consistency and human–AI interaction.

We're building toward AI agents that don't simply respond differently, but become recognisably themselves over time.