[2608.22752] The Compaction Cliff in Long-Running AI Agent Memory

The paper The Compaction Cliff in Long-Running AI Agent Memory identifies a structural failure mode where type-blind hierarchical summarization causes critical safety rules to be disproportionately lost during context compaction. On production agent configurations, safety-rule recall dropped to 53% after one compaction round and fell to 10% after five rounds when using standard compaction prompts on models like Claude Sonnet 4.6.

To mitigate this, the authors propose Knowledge Triage, a framework that classifies agent knowledge into five types and applies specialized retention policies: TypeCompact: Rewrites items in place using per-type fidelity constraints to preserve exact wording for rules. TypeDecompose: Partitions overly large topics into sub-topics, replicating in-scope safety rules across partitions to prevent loss. * TypeRetrieve: Fetches items from external storage while pinning in-scope rules to ensure they remain accessible.

This approach achieves 2–4× higher constraint recall in compaction and 0% locality violations in decomposition compared to uniform strategies, addressing the "Compaction Cliff" where long-running agents silently forget enforceable constraints.

The paper examines a failure mode in long-running AI agent systems that rely on memory compaction—the repeated summarization, pruning, or compression of accumulated context to keep interactions within finite context windows. It identifies what the authors call the compaction cliff: over many agent turns, compaction can preserve superficially coherent summaries while silently discarding task-critical details such as latent constraints, tool-state dependencies, user preferences, error history, or temporal ordering. Performance may remain acceptable for a long period, but once the agent later needs one of these omitted details, behavior can degrade abruptly rather than gradually. The central insight is that the problem is not simply “running out of context,” but rather the accumulation of information loss across multiple compression cycles, especially for details that are unimportant locally but important globally.

A key contribution is a structured account of why compaction is dangerous and how agent memory architectures can be made more robust. The material frames compaction as a state-management problem rather than just a prompt-truncation problem, and it proposes mitigation strategies such as tiered memory separating working context from episodic and semantic stores, importance-aware compression that preserves constraints and open obligations, provenance tracking so summaries can be re-grounded in original evidence, and structured state representations for goals, tools, and task invariants. It also emphasizes evaluation under long-horizon conditions, where the system is tested not only on immediate task completion but on whether it can still recover or reason about details that were compressed many steps earlier.

The work matters because compaction is likely to become a default mechanism for agents that operate over hours, days, or recurring workflows. Without careful memory design, agents may appear reliable in short demos while failing in persistent tasks due to hidden state corruption. The paper therefore provides both a diagnostic lens for debugging long-horizon agent failures and practical architectural guidance for building memory systems that remain auditable, recoverable, and usable as context is repeatedly rewritten.