research(context): Adaptive Focus Memory — three-level fidelity (Full/Compressed/Placeholder) for context budget (arXiv:2511.12712)

[bug-ops]

Description

Adaptive Focus Memory (AFM) assigns each past message one of three fidelity levels—Full, Compressed, or Placeholder—based on semantic relevance, temporal decay, and importance classification, packing messages under a fixed token budget.

This is directly applicable to zeph-context's assembler.rs and manager.rs (changed in CI-799). Current Zeph uses binary keep/discard; AFM's three-level model preserves structural awareness of discarded segments (Placeholder keeps position + ID without tokens), enabling better re-retrieval.

Relevance to Zeph

• crates/zeph-context/src/assembler.rs — message assembly under token budget
• crates/zeph-context/src/manager.rs — context budget management
• crates/zeph-agent-context/src/service.rs — inject_semantic_recall entry point
• Complements: research(context): ACDL — formal language for describing agentic LLM context structure and evolution (arXiv:2605.01920) #3976 (ACDL), research(memory): MEMTIER — tripartite tiered memory with five-signal retrieval and PPO-based weight adaptation (arXiv:2605.03675) #3979 (MEMTIER)

Implementation Sketch

1. Add ContextFidelity { Full, Compressed, Placeholder } enum to context message metadata
2. In assembler: select fidelity per message based on (semantic_distance_to_current, age, importance_score)
3. Placeholder = store only (message_id, role, token_count, importance_score) — zero token cost in prompt
4. On re-access: promote Placeholder → Compressed/Full via memory retrieval

Complexity vs Benefit

• Complexity: Medium | Benefit: High (reduces context tokens 40-60% in long sessions without losing structural history)

Source

• Paper: https://arxiv.org/abs/2511.12712
• Spec: .local/specs/NNN-afm-fidelity-levels/spec.md before implementation