ActiveGraph: Event-Sourced Reactive Graphs for Auditable AI Agents

A recent publication on arXiv (2605.21997) presents ActiveGraph, a runtime designed for agentic systems that reverses the conventional structure by establishing an append-only event log as the foundational truth. Rather than constructing agents around a language model with subsequent logging, ActiveGraph employs a functioning graph that serves as a deterministic representation of the event log. Various behaviors—ranging from standard functions, classes, and LLM-supported routines to logic linked to typed edges—respond to alterations in the graph and generate new events. Coordination is solely managed through the shared graph, eliminating direct instructions between components. This architecture offers three advantages absent in retrieval-and-summarization memory systems: deterministic replay from the log, inexpensive forking at any event without re-executing the shared prefix, and comprehensive lineage from high-level objectives to specific events. The paper can be accessed on arXiv.