The Compaction Cliff in Long-Running AI Agent Memory
Quick summary
arXiv:2608.22752v1 Announce Type: new Abstract: A safety rule and an episodic log compete for the same tokens in an AI agent's context. When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable. On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53\% of safety rules after one compaction round and 10\% after five. We name this the Compaction Cliff. We address it with Knowledge Triage, a framework that classifies each line of an agent's knowledge base by type and routes each type through its
Key takeaways
- arXiv:2608.22752v1 Announce Type: new Abstract: A safety rule and an episodic log compete for the same tokens in an AI agent's context.
- When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable.
- On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53\% of safety rules after one compaction round and 10\% after five.
Why it matters
“The Compaction Cliff in Long-Running AI Agent Memory” shows why AI risk cannot be reduced to answer accuracy. Access controls, logging, human approval and incident response need to be designed into the workflow from the start.
Member comments