AI Agents Need Memory Control Over More Context

TL;DR

The paper tackles the reliability challenges of long-horizon AI agents caused by unbounded memory growth, drift, and hallucination. It introduces the Agent Cognitive Compressor (ACC), a memory controller that maintains a single, bounded internal state called the Compressed Cognitive State (CCS) and separates external artifact recall from state commitment via a schema-constrained compressor. Through a judge-driven, live evaluation framework across IT operations, cybersecurity, healthcare, and finance, ACC shows consistent memory bounding, reduced hallucination and drift, and improved multi-turn stability compared to transcript replay and retrieval-based approaches. The work demonstrates that memory governance is a practical foundation for reliable long-horizon AI agents and outlines directions for further validation, task-adaptive CCS schemas, and multi-agent extensions.

Abstract

AI agents are increasingly used in long, multi-turn workflows in both research and enterprise settings. As interactions grow, agent behavior often degrades due to loss of constraint focus, error accumulation, and memory-induced drift. This problem is especially visible in real-world deployments where context evolves, distractions are introduced, and decisions must remain consistent over time. A common practice is to equip agents with persistent memory through transcript replay or retrieval-based mechanisms. While convenient, these approaches introduce unbounded context growth and are vulnerable to noisy recall and memory poisoning, leading to unstable behavior and increased drift. In this work, we introduce the Agent Cognitive Compressor (ACC), a bio-inspired memory controller that replaces transcript replay with a bounded internal state updated online at each turn. ACC separates artifact recall from state commitment, enabling stable conditioning while preventing unverified content from becoming persistent memory. We evaluate ACC using an agent-judge-driven live evaluation framework that measures both task outcomes and memory-driven anomalies across extended interactions. Across scenarios spanning IT operations, cybersecurity response, and healthcare workflows, ACC consistently maintains bounded memory and exhibits more stable multi-turn behavior, with significantly lower hallucination and drift than transcript replay and retrieval-based agents. These results show that cognitive compression provides a practical and effective foundation for reliable memory control in long-horizon AI agents.

Figures (7)

• Figure 1: Agent architecture incorporating the Agent Cognitive Compressor (ACC). ACC operates as a cognitive memory controller that constructs and commits a bounded Compressed Cognitive State (CCS) via a schema constrained Cognitive Compressor Model (CCM). CCS is the sole persistent internal state maintained across turns and conditions downstream reasoning, tool use, and action, while ACC remains decoupled from policy execution and environment interaction.
• Figure 2: ACC state commitment mechanism for producing the next Compressed Cognitive State $\mathrm{CCS}_t$ under the schema constraint $\mathcal{S}_{\mathrm{CCS}}$, using the current interaction $x_t$, the previously committed state $\mathrm{CCS}_{t-1}$, and the qualified recalled set $A_t^{+}$.
• Figure 3: Example Compressed Cognitive State (CCS) for an IT operations task, serialized using a TOON style token oriented representation.
• Figure 4: ACC integration patterns in agent architectures: ReAct multi-turn tool execution (left) and plan execute reflect control (right).
• Figure 5: Memory footprint across domains over 50-turn live testing. Baseline grows with turn count, while Retrieval (top-3 artifacts) and ACC remain bounded.