Long term memory cortex for agents that maintains tensions (github.com)

Daftari introduces a transformative long-term memory cortex for large language model (LLM) agents, emphasizing persistent memory as a crucial tool for knowledge accumulation and reasoning. Unlike existing models that treat memory as a stateless feature, Daftari employs a structured markdown vault where agents can read, write, and curate data over time. This allows for durable, portable memory that is owner-controlled and enhances the model's ability to handle complex tasks by preserving contradictions and open tensions rather than flattening them into false certainties. Significantly, Daftari's architecture facilitates an accumulation process, where each entry builds on prior ones, creating a richer context for agents to draw from. It incorporates a robust provenance tracking system that maintains the integrity of information, while also allowing concurrent access and safe mutations through file-level locking and version control. The set of features—such as tension logging, role-based access controls (RBAC), and a flexible integration with existing markdown systems—ensures that agents can not only compile responses efficiently but also reason effectively with the historical context of information, which is vital for future AI applications in dynamic, knowledge-intensive environments.