Title: DMF: A Deterministic Memory Framework for Conversational AI Agents

Abstract

To function effectively over extended interaction periods, conversational AI agents necessitate memory architectures that balance semantic coherence with scalability. Current methodologies largely depend on Large Language Model (LLM)-based summarization during the writing phase; however, this reliance introduces significant drawbacks, including non-deterministic outcomes, rising token expenditures, and a lack of transparency in how data is pruned. In response, we introduce the Deterministic Memory Framework (DMF), a CPU-centric solution that discards generative memory compression in favor of a strictly deterministic pipeline. This approach leverages classical NLP analysis, vector geometry, and mathematical scoring.

Within DMF, every conversational turn is evaluated using a Survival Score ($\Omega$). This metric is derived from deterministic content indicators, conversational signals, and structured provenance, which are integrated via a logistic projection. To manage relevance as new exchanges occur, the framework employs an interaction-count decay law, expressed as $\Omega_{\mathrm{eff}}(\Delta n)$. Here, $\Delta n$ represents the volume of subsequent interactions rather than elapsed time, thereby maintaining complete determinism throughout the process.

This paper details the mathematical underpinnings of DMF, its structured recall mechanism, the protocol for pruning decisions, and the associated evaluation standards. Our experimental validation utilizes a custom benchmark built upon the LoCoMo and LongMemEval datasets. We benchmark DMF against Mem0, a widely adopted memory layer for AI agents. The results demonstrate that DMF matches Mem0 in accuracy while requiring zero tokens to prepare memory contexts and reducing overall token usage by a factor of 5 to 242 across the entire conversation. These findings confirm that LLM calls can be entirely removed from the memory management loop, effectively driving token costs to near zero and establishing a viable path toward deterministic memory systems for conversational AI.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC