RhizoDB: A Bounded Activation-Flow Architecture for Graph-Based Memory Systems

TL;DR

RhizoDB is a graph-based memory architecture for agentic systems that maintains memory stability through bounded activation flow, stability-modulated temporal decay, depth-limited propagation, periodic washout, and structural pruning.

Problem

In agentic systems, repeated signal composition and memory accumulation can lead to activation runaway or forgetting imbalance. Existing memory systems struggle to simultaneously ensure long-term dependency and stability.

Approach

RhizoDB operates as a nonlinear dynamical system over a directed graph. Each node has an activation value that decays over time modulated by a stability coefficient. Activation propagation is depth-limited, and periodic washout removes low-activation nodes, along with structural pruning.

Results & Contribution

This technical report provides a mathematical formulation of RhizoDB and relates its local activation-control mechanisms to broader stabilization approaches for repeated signal composition. No experimental results are presented; the focus is on conceptual contribution.