research(context): AgeMem — RL-trained proactive summarization before context fills (arXiv:2601.01885)

[bug-ops]

Description

AgeMem (Agentic Memory) trains memory operations—store, retrieve, update, summarize, discard—as callable tools via a three-stage RL pipeline with step-wise GRPO. Crucially, learned policies discover proactive summarization: the agent summarizes intermediate results before the context window fills, not after hitting the limit.

Directly maps to CI-799's zeph-context and zeph-agent-context changes. Current Zeph compaction triggers reactively (budget threshold); AgeMem demonstrates a learned, proactive trigger is significantly more effective.

Relevance to Zeph

• zeph-context: manager.rs, assembler.rs — proactive compaction trigger policy
• zeph-agent-context: summarization/scheduling.rs — could replace heuristic scheduling with RL-informed scheduling
• Related open issue: research(context): ACDL — formal language for describing agentic LLM context structure and evolution (arXiv:2605.01920) #3976 (ACDL formal notation), research(memory): MEMTIER — tripartite tiered memory with five-signal retrieval and PPO-based weight adaptation (arXiv:2605.03675) #3979 (MEMTIER tiered retrieval)

Implementation Sketch

1. Expose compaction/summarization trigger as a callable tool available to the agent
2. Collect trajectory data (context size, task progress, upcoming tool calls)
3. Train a lightweight GRPO policy to predict optimal summarization moments
4. Alternatively: use a simple learned threshold on context_budget_used × task_horizon

Complexity vs Benefit

• Complexity: High (RL training pipeline) | Benefit: High (proactive compaction prevents context blowout mid-task)
• Near-term: expose trigger as tool (low complexity), defer RL training to post-v1.0

Source

• Paper: https://arxiv.org/abs/2601.01885
• Spec: create .local/specs/NNN-agemem-proactive-compaction/spec.md before implementation