research(memory): ACON agent context compression — 26-54% peak token reduction via guideline optimization (arXiv:2510.00615) #2433
Description
Summary
arXiv:2510.00615 — ACON: Optimizing Context Compression for Long-horizon LLM Agents
Technique
ACON (Agent Context Optimization) compresses both environment observations and interaction histories into concise yet informative condensations, using compression guideline optimization in natural language space. The guidelines are themselves learned (not hand-crafted) — the system improves its compression strategy over time.
Results: 26-54% peak token reduction while largely preserving task performance.
Applicability to Zeph
Zeph's compress_context uses fixed summarization prompts. ACON suggests:
- Treating the summarization prompt as a learnable policy (compression guideline)
- Optimizing the guideline on past sessions (what information turned out to be needed after compaction vs. discarded)
- Separate compression strategies for: tool outputs, assistant reasoning, user context
Could be applied to compaction_provider path in zeph-core compression, potentially improving the quality of compacted context with fewer tokens.
Priority
P2 — directly applicable to zeph-core compaction quality improvement.
Source
arXiv:2510.00615 (accepted ICLR 2026)