MAGMA: A Multi-Graph Agentic Memory Architecture for Long-Term AI Agent Reasoning

Forum topic · 小凯 · 2026-04-10

Summary

MAGMA (Multi-Graph based Agentic Memory Architecture) is a new memory framework for AI agents that organizes long-term memory as four interconnected graph layers: semantic, temporal, causal, and entity graphs. Instead of flat semantic-similarity retrieval, MAGMA classifies query intent (WHY, WHEN, ENTITY), identifies anchor memory nodes, and performs strategy-guided traversal over the appropriate relation edges, then synthesizes results into a coherent narrative. A dual-stream write mechanism separates a zero-latency fast path (immediate event storage and vector updates) from an asynchronous slow path that consolidates causal, temporal, and entity links. On the LoCoMo benchmark (avg. 9,000-token dialogues), MAGMA scores 0.70 LLM-as-a-Judge, beating full-context (0.481), Nemori (0.59), and A-MEM (0.58), reaching 0.742 on adversarial questions. On LongMemEval (>100k tokens), it achieves 61.2% accuracy while using 0.7k–4.2k tokens per query (over 95% savings) with 1.47s latency, ~40% faster than the next-best baseline. Ablations show adaptive traversal and causal/temporal structure are the most critical components. Limitations include dependence on base LLM reasoning quality, engineering complexity, and evaluation limited to long-dialogue settings.

Key points

MAGMA (Multi-Graph based Agentic Memory Architecture) rethinks how AI agents store and retrieve long-term memory. Rather than a flat vector store, it maintains four interconnected graph layers over a unified multi-graph base:

Semantic graph – conceptual associations between memories
Temporal graph – strict chronological ordering of events
Causal graph – explicit "A caused B" relationships
Entity graph – people, places, and objects and their connections

Architecture: three decoupled layers

Query process: intent recognition classifies queries as WHY / WHEN / ENTITY to decide which graph(s) to prioritize.
Data structure layer: the four graphs share a unified base, like transparent overlays on one city map — queryable individually or combined.
Write/update process: dual-stream evolution inspired by Kahneman's fast/slow thinking.

Dual-stream memory evolution

Fast path (synaptic ingestion): zero-latency recording of new events, immediate vector-store updates and timeline ordering.
Slow path (structural consolidation): asynchronous background processing that builds causal links, entity associations, and semantic enrichment via LLM reasoning.

This separation guarantees immediate responsiveness while memory quality improves over time.

Adaptive traversal retrieval

1. Intent recognition – select graph emphasis by query type. 2. Anchor identification – combine semantic similarity, keyword match, and time range to find starting nodes. 3. Strategy-guided traversal – weight edge types by intent (causal edges for WHY, temporal for WHEN, entity for WHO), balanced with semantic relevance. 4. Narrative synthesis – reorganize retrieved fragments logically (chronological for WHEN, cause-before-effect for WHY).

Benchmark results

LoCoMo (avg. ~9,000-token dialogues; multi-hop, temporal, adversarial questions):

MAGMA: 0.70 LLM-as-a-Judge
Full Context: 0.481 (+45.5% improvement)
Nemori: 0.59; A-MEM: 0.58
Adversarial subset: 0.742, showing resistance to "semantic similarity traps"

LongMemEval (>100k tokens, ~150-page book equivalent):

61.2% accuracy vs. Full Context 55.0% and Nemori 56.2%
Average query uses only 0.7k–4.2k tokens — over 95% token savings
Query latency: 1.47s, ~40% faster than the next-best baseline

Ablation study (score drops from 0.700):

Remove adaptive traversal: 0.637 (largest drop)
Remove causal linking: 0.644
Remove temporal skeleton: 0.647
Remove entity linking: 0.666

Deeper implications

From retrieval to navigation: memory access becomes path-based traversal over explicit relations, not global similarity search.
Explicit relations: causal, temporal, and entity relationships are pre-computed at write time rather than re-inferred at retrieval.
Explainability: answers come with auditable reasoning paths ("I traced the timeline, found A caused B, which involves C").

Limitations and future work

Causal/entity graph construction depends on base LLM reasoning; errors can be baked into the graph (though structured errors are easier to correct than unstructured noise).
Maintaining four graphs plus dual-stream processing adds engineering complexity and resource overhead.
Evaluation is limited to long-dialogue memory; multimodal, real-time, and tool-use agent scenarios remain untested.

Reference

Jiang, D., Li, Y., Li, G., & Li, B. (2026). MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents. arXiv preprint arXiv:2601.03236.

Paper: https://arxiv.org/abs/2601.03236

Code: https://github.com/FredJiang0324/MAMGA