research(context): ARC — active + reflection-driven context management for long-horizon agents (arXiv:2601.12030)

[bug-ops]

Description

ARC (Active and Reflection-driven Context Management) gives the LLM explicit agency over its own context: the agent can proactively request summarization if it judges the context unwieldy. ARC is the only agent in a broad corpus survey that delegates the compaction decision to the LLM itself (vs. fixed heuristics).

Complements CI-799's zeph-agent-context changes. Current Zeph compaction is scheduler-driven; ARC's model would let the agent signal 'I need a summary now' mid-task.

Relevance to Zeph

• crates/zeph-agent-context/src/summarization/scheduling.rs — add LLM-initiated compaction signal
• crates/zeph-core/src/agent/mod.rs — agent loop processes self-compaction request like a tool call
• Related: research(context): AgeMem — RL-trained proactive summarization before context fills (arXiv:2601.01885) #4016 (AgeMem proactive compaction), research(context): ACDL — formal language for describing agentic LLM context structure and evolution (arXiv:2605.01920) #3976 (ACDL)

Implementation Sketch

1. Add request_compaction(reason: String) as an internal tool available to the agent
2. In agent loop: if tool_call.name == 'request_compaction', trigger immediate summarization and retry
3. Track compaction-request frequency in metrics for future RL training signal

Complexity vs Benefit

• Complexity: Low-Medium (expose as tool, integrate in loop) | Benefit: Medium (agent-initiated compaction is more accurate than heuristic scheduling)

Source

• Paper: https://arxiv.org/abs/2601.12030