feat: Memory contamination detection for agent memory debugging (State Contamination)

[acailic]

Paper Reference

• Title: State Contamination in Memory-Augmented LLM Agents
• Authors: Yian Wang, Agam Goyal, Yuen Chen, Hari Sundaram
• Year: 2026
• URL: https://arxiv.org/abs/2605.16746
• Venue: arXiv preprint

Paper Summary

Studies "memory laundering" where adversarial context is compressed into seemingly-safe memory summaries that still influence behavior downstream. Introduces sub-threshold propagation gap (SPG) metric for measuring hidden influence in compressed memories.

Proposed Feature

Implement memory state validation and contamination detection:

Core Capabilities

• Memory Influence Tracking: Track how past experiences influence current agent decisions through memory
• Contamination Detection: Detect when adversarial or corrupted information persists through memory compression
• SPG Metric Dashboard: Show the sub-threshold propagation gap for memory states
• Memory Sanitization: Flag and offer to sanitize memory entries with detected contamination

Technical Approach

• Add memory state tracking to the SDK's session capture
• Implement influence propagation analysis between memory entries and decisions
• Build memory health dashboard showing contamination scores per memory entry
• Add sanitization controls in the UI

Impact

Critical for agents using RAG, long-term memory, or compressed context. Memory bugs are extremely difficult to debug manually — this provides automated detection.

Labels

enhancement, paper-inspired, safety