[ 06 ] — Research

A two-timescale architecture,
published as a preprint.

MindForm is built on StableMind, a framework we describe in a preprint on Zenodo: model-agnostic by design and open-source at the core.

Preprint · 2026
MindForm · Technical paper

StableMind: A Two-Timescale Architecture for Persistent Personality in LLM Agents

Hasan Mavlonov, ECUST · Major in Artificial Intelligence

We propose StableMind, a memory- and reflection-based architecture that decouples short-term belief plasticity from long-term trait evolution. Trait evolution occurs only when drift satisfies tri-fold criteria of magnitude, confidence, and observation count — ensuring updates are bounded and stepwise.

Endorsed for arXiv submission · Nuo Chen, NUS — author of "Diversity Collapse in Multi-Agent LLM Systems"
Selected citations
[1] Li et al. — A persona-based neural conversation model. ArXiv 1603.06155
[2] Packer et al. — MemGPT: Towards LLMs as operating systems. ArXiv 2310.08560
[3] Park et al. — Generative agents: Interactive simulacra of human behavior. UIST '23
[4] Shinn et al. — Reflexion: language agents with verbal reinforcement learning. NeurIPS '23
Benchmarks

Benchmark results coming soon.

Milestones
  1. Jun 2026

    StableMind preprint published → (opens in a new tab)

    Zenodo · DOI 10.5281/zenodo.20827145

  2. Aug to Oct 2026

    10 node white papers published

Also
Read the preprint → View on GitHub → Back to the homepage