Multi-Turn Agentic Context Decay & State Pollution Benchmark

📌 One-Sentence Summary

This benchmark measures the decay of agent memory accuracy over 50 interaction steps， revealing that while flat RAG and linear KV-cache models suffer significant state pollution (dropping to ~60% recall)， the XORAS Poincaré Memory system maintains near-perfect recall at 99.8% with zero pollution.

📝 Summary

The experiment tests four different memory architectures for their ability to retain context over a long conversation. The results show that traditional methods like flat Euclidean RAG and linear KV-cache (Gemma-4) degrade significantly， with accuracy falling to around 60% and high state pollution rates. In contrast， the XORAS 18，432-D Poincaré Memory demonstrates exceptional stability， maintaining over 99% accuracy and zero pollution throughout the 50 steps.

💡 Main Points

Context Decay in Traditional Models

Flat RAG and linear KV-cache models exhibit significant context decay， with recall dropping from an initial 97.2% to approximately 60% over 50 steps. This indicates that these architectures struggle to maintain distinct memory states， leading to 'topic bleed' and state pollution.

XORAS Poincaré Memory Performance

The XORAS system， utilizing a 18，432-dimensional Poincaré embedding， outperforms the competition by maintaining a recall rate of 99.8% and zero state pollution. This suggests that the hyperbolic geometry of the Poincaré space is effective at isolating memory states.

Benchmark Methodology

The benchmark uses a 'BBQ' (Binary BBQ) verification protocol to assess the fidelity of the agent's memory. The high throughput of 3.16M ops/s and the use of Z3 SMT proofs ensure that the results are robust and mathematically verified.

💬 Key Quotes

Linear KV-cache and flat RAG suffer topic bleed， dropping recall from 97.2% to 60.8%.

📊 Article Meta

AI Screening： 90

Source： DEV Community: machinelearning

Author： Anthony Oxendine

Category： 人工智能

Language： 英文

Read Time： 1 min

Word Count： 109

Tags：

AI 与智能应用 , AI Agent , RAG / 检索增强 , 模型训练与推理 , Agent记忆