Short-term memory
Fast adaptation layer for immediate user context, goals, and conversational state. Optimized for responsiveness and local relevance.
Early-stage AI startup
Building the personality layer for language model agents.
Today’s AI systems are highly capable but behaviorally unstable: they shift identity with prompt context and fail to preserve coherent character over time. Mind Form is developing an explicit two-timescale architecture that separates short-term adaptation from long-term personality persistence.
Problem: AI systems lack persistent identity
Most LLM agents are reactive. They adapt to local context, but the adaptation dominates the model’s behavior, producing drift across sessions and inconsistent decisions under slight framing changes.
This limits their reliability for collaborative research, knowledge work, and long-lived user relationships.
Observed failure modes
- Session-to-session personality drift
- Overfitting to local prompt style
- No principled identity memory model
- Weak continuity in long-horizon tasks
Solution: Two-timescale architecture
Long-term personality
Slow-moving identity layer that encodes stable traits, preferences, behavioral constraints, and policy consistency across interactions.
The objective is not static behavior. The objective is coherent evolution: agents that can learn without losing identity.
Mission
Build foundational methods for persistent AI personality systems and deploy them as infrastructure for next-generation agents. We prioritize technical rigor, empirical evaluation, and research transparency.
We are recruiting students, researchers and engineers interested in agent memory, behavioral consistency, cognitive modeling, and alignment-adjacent systems design.
Join Team