Source

ACON: Optimizing Context Compression for Long-horizon LLM Agents (ICLR 2026 submission)

Finding

ACON optimizes compaction by learning from failures: when full context succeeds but compressed context fails on the same task, an LLM analyzes the failure and updates the compression guidelines in natural language. Gradient-free — works with closed-source models. Achieves 26–54% token reduction with up to 46% performance improvement for smaller LMs.

Applicability

Zeph's compaction pipeline (apply_deferred_summaries, chunked_compaction in zeph-memory) uses a fixed LLM prompt. ACON's failure-driven approach could be layered on top:

1. Store a "compression guidelines" document (initially empty)
2. When a post-compaction session produces a task failure attributable to missing context, log the pair (compressed_context, failure_reason)
3. Periodically run an LLM to update the guidelines from failure pairs
4. Inject the current guidelines into the summarization prompt

Related: #1607 (anchored summarization), #1609 (task-continuation metric — a natural failure signal for this loop)

Priority

High — directly improves the known multi-session context retention problem; no training infrastructure needed.