Paper

Title: D-MEM: Dopamine-Gated Agentic Memory via Reward Prediction Error Routing
arXiv: https://arxiv.org/abs/2603.14597
Published: 2026-03-15

Key Technique

Biologically-inspired Reward Prediction Error (RPE) signal gates memory evolution:

• Low-RPE (routine/repeated) inputs → O(1) fast buffer, bypasses full evolution
• High-RPE (factual contradictions, preference shifts) → dopamine signal → O(N) full graph evolution pipeline

Lightweight Critic Router evaluates Surprise + Utility. Decouples short-term interaction from cognitive restructuring.

Results: 80% token reduction, eliminates O(N²) bottleneck, outperforms baselines on multi-hop reasoning + adversarial resilience.

Why Relevant to Zeph

Zeph runs full MAGMA graph extraction on every memory_save call. D-MEM's approach would:

1. Route low-surprise turns (repetitive queries, confirmations, small talk) to fast SQLite write without full LLM graph extraction
2. Reserve expensive MAGMA extraction for high-RPE turns (new entity introductions, belief contradictions, goal changes)
3. Directly reduces LLM call volume for graph extraction — relevant given extraction_provider cost

Integration point: A-MAC (memory admission control at write-gate) + MAGMA extraction threshold. The RPE signal extends A-MAC from binary write/skip to tiered: write-to-sqlite / write-to-sqlite+graph.

Complements #2416 (RL-based admission) but focused on graph pipeline cost rather than memory write decision.

Priority Rationale

P2: practical cost reduction (80% token savings) for an existing high-cost path in Zeph. Addresses observable overhead from per-turn MAGMA extraction in long sessions.