Most agent frameworks follow a familiar pattern: conversation loop first, tools added, rules layered on top, and finally a logging layer bolted on for observability. Yohei Nakajima presents Active Graph, a runtime system that completely inverts this arrangement by making the append-only event log the foundational source of truth rather than an afterthought.

In Active Graph, the working graph emerges as a deterministic projection of the event log itself. Behaviors—whether ordinary functions, classes, LLM-backed routines, or logic attached to typed edges—react to changes in this shared graph and emit new events. Coordination happens entirely through the graph structure, eliminating direct component-to-component instructions. This architectural choice, submitted on May 21, 2026, fundamentally changes how agentic systems maintain state and process information.

The design yields three critical properties that traditional memory systems lack. First, deterministic replay of any execution run becomes possible directly from its log. Second, forking operations branch runs at any event point without re-executing shared prefixes. Third, complete end-to-end lineage traces high-level goals down to individual model calls that produced each artifact.

The paper connects this work to the Baby AGI lineage and prior graph-memory research, suggesting the substrate suits self-improving agents particularly well. A worked diligence example demonstrates reconstructing full causal structures solely from log data, though the authors acknowledge they do not fully demonstrate all claimed benefits.