Source

arXiv:2601.07190 — Active Context Compression: Autonomous Memory Management in LLM Agents (Jan 2026)

Summary

Focus: agent autonomously decides when to consolidate history into a persistent "Knowledge" block and prune raw history. Achieves 22.7% token reduction (up to 57% on individual tasks) at no accuracy cost by treating the agent itself as the compression decision-maker.

Key Ideas

• compress_context tool: callable by the agent itself (not threshold-triggered)
• Knowledge block: a persistent, named first-class context region separate from raw history
• Exploration-driven trigger: compress when exploration stalls, not on a fixed token budget — "slime mold" heuristic
• 6 autonomous compressions per task on average (useful calibration target for testing)

Applicability to Zeph

Directly applicable to zeph-memory compaction and context assembly:

• Zeph uses passive threshold-triggered compaction (compaction_threshold=0.80, soft_compaction_threshold)
• Focus proposes replacing this with an agent-invokable compress_context tool
• The Knowledge block maps to a new named region in ContextBuilder (alongside system prompt, tools, conversation)
• The trigger heuristic replaces the current token-ratio logic with a trajectory-based approach

Implementation Sketch

1. Add compress_context to native tool set (calls existing compaction pipeline)
2. Add a "Knowledge" context region in ContextBuilder that survives compaction
3. Add token trajectory tracking to decide when compaction makes sense (beyond simple threshold)
4. Expose via config: [memory.compression] strategy = "autonomous"

Complexity

Low-Medium — Knowledge block and compress tool are contained changes; trigger heuristic is the main design decision.