Standard Retrieval-Augmented Generation (RAG) treats vector databases as passive dictionaries. This technical report introduces a structural identity layer for LLM agents where the retrieval space is dynamic and reshaped by execution feedback.
Core Contributions:
Divergence-Driven Navigation: When an agent tests a hypothesis (a "pivot") and it fails, the system does not search for similar facts. Instead, it flips the vector search space (ORDER BY ... DESC) constrained by domain overlap, forcing the agent to explore dissimilar hypotheses and preventing repetitive reasoning loops.
Model-Agnostic Identity: Agent state (knowledge, learned faults, execution context) is persisted as contrast embeddings in pgvector, completely decoupled from the executor model. We provide empirical evidence of seamless intra-session model handoffs (e.g., DeepSeek for planning → GLM-5.2 for execution → DeepSeek for synthesis) where task continuity is maintained by the structural layer, not model memory.
Zero-Cost Behavioral Supervisor: A lightweight local 4B parameter model enforces the use of the structural navigation layer over speculative reasoning, consuming zero tokens from the frontier models and adding negligible latency (0.1ms).
The document includes the complete relational schema, the core SQL/Python logic for the divergence mechanism, empirical logs of automated divergent exploration, and a transparent discussion of current system limitations.
https://zenodo.org/records/20837421
Standard Retrieval-Augmented Generation (RAG) treats vector databases as passive dictionaries. This technical report introduces a structural identity layer for LLM agents where the retrieval space is dynamic and reshaped by execution feedback.
Core Contributions:
The document includes the complete relational schema, the core SQL/Python logic for the divergence mechanism, empirical logs of automated divergent exploration, and a transparent discussion of current system limitations.
https://zenodo.org/records/20837421