The Compaction Cliff in Long-Running AI Agent Memory

The Compaction Cliff is a structural failure mode in long-running AI agent memory where type-blind compaction causes safety rules to degrade rapidly. Published at CIKM 2026 by Saber Zerhoudi, Jelena Mitrovic, and Michael Granitzer, the research identifies that when agents summarize their context, safety constraints compete with episodic logs for token space, leading to significant loss of enforceable rules.

Key findings include: Severity: On production configurations using Claude Code’s /compact prompt on Sonnet 4.6, safety rule recall drops from 53% after one compaction round to just 10% after five rounds. Mechanism: Standard hierarchical summarization treats all context uniformly, paraphrasing or dropping precise safety wording that requires exact retention. * Solution: The authors propose Knowledge Triage, a framework that classifies knowledge into five types and applies distinct retention policies, using operators like TypeCompact (per-type fidelity) and TypeRetrieve (pinned in-scope rules) to prevent this degradation.

The paper investigates a failure mode that becomes especially visible in long-running AI agents: as conversations, task histories, and accumulated context grow, systems often rely on compaction—summarizing, pruning, or rewriting prior context to fit fixed model windows. Rather than treating compaction as a neutral compression step, the work frames it as a critical reliability boundary. It examines how aggressive or poorly structured compaction can cause abrupt performance degradation, a “compaction cliff,” in which an agent that previously had the necessary information loses the ability to use it effectively. The central concern is not merely that some details are forgotten, but that compaction can destroy the latent state needed for coherent continuation: unresolved constraints, implicit assumptions, tool-call provenance, cross-session dependencies, and the distinction between stable facts and transient working context.

Its key contribution is to make this degradation explicit and analyzable. The material connects memory management to long-horizon task performance, showing that compaction quality is a first-class design problem rather than an implementation detail. It likely emphasizes that different kinds of information have different fragility under summarization: high-level goals may survive, while low-level invariants, error conditions, user preferences, and intermediate reasoning traces may not. From this, the paper motivates more structured approaches to agent memory, such as layered representations that separate episodic history from semantic knowledge, preserving provenance for important actions, maintaining explicit task state, and using incremental or verified compaction instead of wholesale context rewriting.

This matters because long-running agents are moving beyond single-turn assistance into persistent workflows, personalization, and multi-day task execution. In such settings, compaction is unavoidable, but its effects are often silent: the agent may still appear fluent while losing the operational context needed to act correctly. By identifying the compaction cliff and the information-loss mechanisms behind it, the work provides a useful lens for evaluating memory architectures, designing safer context-management pipelines, and building benchmarks that test not just recall but sustained task competence after context reduction.