Adaptive Memory Forgetting: How AI Agents Can Balance Relevance and Efficiency in Long Conversations

The Forgetting Problem in AI Agents

Long-horizon conversational agents need persistent memory, yet uncontrolled accumulation causes temporal decay and false memory propagation. A new paper introduces an adaptive budgeted forgetting framework.

The Problem

Benchmarks show severe degradation with persistent memory:

• LOCOMO/LOCCO: performance drops from 0.455 to 0.05 across stages
• MultiWOZ: 78.2% accuracy with 6.8% false memory rate

The Solution: Adaptive Budgeted Forgetting

The framework regulates memory through relevance-guided scoring and bounded optimization, integrating:

• Recency — How recently was the memory accessed?
• Frequency — How often is it referenced?
• Semantic alignment — Does it relate to the current context?

Results

• Improved long-horizon F1 beyond 0.583 baseline
• Higher retention consistency
• Reduced false memory behavior
• No increase in context usage

Why It Matters

As AI agents handle longer and more complex conversations, naive memory accumulation becomes a liability rather than an asset. Structured forgetting preserves reasoning performance while preventing unbounded memory growth.

arXiv: 2604.02280