Source

arXiv:2601.03236 — "MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents"

Summary

Organizes memory across four explicit relational graphs (semantic, temporal, causal, entity) with policy-guided traversal. Dual-stream architecture separates fast ingestion from async structural consolidation. Outperforms flat semantic search on LoCoMo and LongMemEval benchmarks.

Applicability to Zeph

HIGH — zeph-memory graph memory subsystem.

Zeph already has graph memory with entity resolution and BFS/FTS5 retrieval. MAGMA's decomposition directly extends the existing graph_edges schema:

• Add edge_type variants: semantic, temporal, causal (currently only entity co-occurrence)
• Policy-guided traversal maps onto spreading activation ([memory.graph.spreading_activation])
• Dual-stream pattern matches Zeph's async graph extraction architecture

Implementation Direction

• Extend graph_edges.edge_type to include causal/temporal types
• Update GraphExtractor LLM prompt to extract edge types
• Add causal-aware traversal path to BFS retrieval
• Measure recall improvement on cross-turn fact retrieval scenarios

Priority: P2
Discovered: CI-211 research scan (2026-03-27)