OPEN LINK ↗
// 2h agoRESEARCH PAPER
The Compaction Cliff Exposes Agent-Memory Losses
The paper measures how context compaction degrades safety rules across 20 production agent configurations, finding Claude Code’s Sonnet 4.6 /compact preserves just 53% after one round and 10% after five. It introduces Knowledge Triage, which applies type-specific retention policies to protect safety-critical instructions.
// ANALYSIS
Context compaction is a safety boundary, not merely a token-saving feature; uniformly summarizing rules and episodic history quietly turns memory loss into policy failure.
- –Knowledge Triage classifies agent knowledge and applies separate compact, decompose, and retrieve operations by type.
- –TypeCompact preserves 2–4× more safety rules than competing single-shot compactors, with 96% recall after five rounds.
- –TypeRetrieve reaches 100% recall@50 by prioritizing applicable rules ahead of ordinary relevance.
- –The practical takeaway is clear: AGENTS.md and CLAUDE.md rules need protected, verifiable retention rather than ordinary summarization.
- –The released AgentArtifactCorpus contains 396,934 artifacts from 54,628 public GitHub repositories, giving developers a useful benchmark for memory systems.
// TAGS
the-compaction-cliffagent-memorycontext-engineeringsafetyevaluationresearch
DISCOVERED
2h ago
2026-08-25
PUBLISHED
2h ago
2026-08-25
RELEVANCE
9/ 10
AUTHOR
omarsar0