A new arXiv paper argues that persona-based AI agents need more than a short profile prompt if they are expected to behave consistently over time.
The authors introduce MemoryForge, a framework for synthesizing lifelong memory from a brief target persona. Instead of conditioning a frozen large language model only with static descriptions, the approach builds an autobiographical memory base that the model can retrieve from when responding to situations.
The paper frames this as “memory-based conditioning.” Its components include a context generator for socio-historical grounding, a life organizer for developmental coherence and a multi-resolution simulator that balances broad life structure with detailed memories. The goal is to make role-play, user simulation and other agentic applications less generic.
The limitation is that synthetic memories are still constructed artifacts, not real experience. The work is most relevant as a research direction for long-running agents: if systems are expected to act like persistent characters or users, developers need ways to test whether memory improves coherence without creating misleading claims of human-like understanding.