[Paper] MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents

Source: arXiv - 2601.03236v1

Overview

The paper introduces MAGMA, a new memory architecture for AI agents that moves beyond the traditional “single‑bucket” external memory used in many Retrieval‑Augmented Generation (RAG) systems. By organizing memories as multiple, orthogonal graphs—semantic, temporal, causal, and entity—MAGMA lets an agent traverse the right relationships for each query, yielding more transparent and accurate long‑context reasoning.

Key Contributions

• Multi‑graph memory representation: Each stored fact is simultaneously a node in four distinct graphs (semantic similarity, chronological order, causal links, and entity co‑occurrence).
• Policy‑guided retrieval: Retrieval is cast as a reinforcement‑learning‑style traversal policy that decides which graph edges to follow, making the search adaptive to the query’s intent.
• Agentic memory abstraction: Decouples what is stored from how it is retrieved, enabling fine‑grained control and interpretability of the reasoning path.
• Empirical gains: On the LoCoMo and LongMemEval benchmarks, MAGMA outperforms prior agentic memory systems by 4–9 % absolute accuracy on long‑horizon reasoning tasks.
• Open‑source implementation: The authors release code and pre‑trained graph encoders, facilitating reproducibility and downstream experimentation.

Methodology

1. Memory Encoding

   • When a new piece of information arrives (e.g., a dialogue turn or a knowledge snippet), it is embedded once and then inserted as a node into four separate graphs:
     • Semantic graph – edges based on cosine similarity of embeddings.
     • Temporal graph – directed edges linking newer items to older ones.
     • Causal graph – edges inferred from explicit cause‑effect statements or learned via a causal classifier.
     • Entity graph – edges connecting items that share named entities.

2. Policy‑Guided Traversal

   • Given a user query, a lightweight policy network predicts a sequence of graph‑type selections (e.g., “start with semantic, then follow temporal”).
   • At each step, the policy expands the frontier by traversing edges of the chosen graph, scoring candidate nodes with a relevance model.
   • The traversal stops after a budgeted number of hops or when a confidence threshold is met, producing a ranked list of memory items.

3. Context Construction & Generation

   • The retrieved items are concatenated (or hierarchically structured) and fed to a large language model (LLM) as augmented context.
   • Because the retrieval path is explicit, the system can also surface the graph walk as a “reasoning trace” for debugging or user explanation.

Results & Findings

• Higher accuracy stems from the ability to fetch temporally or causally relevant facts that a pure semantic similarity search would miss.
• Interpretability: The authors show case studies where the retrieved graph walk aligns with human logical steps, something monolithic memories cannot expose.
• Efficiency: Despite maintaining four graphs, the traversal budget stays low (≈ 5 hops on average), keeping latency comparable to standard RAG pipelines.

Practical Implications

• Developer‑friendly debugging – The explicit traversal trace can be logged or visualized, helping engineers pinpoint why a model answered incorrectly.
• Fine‑grained control – Teams can bias the policy toward certain graphs (e.g., prioritize causal links for troubleshooting bots) without retraining the whole LLM.
• Scalable long‑term agents – Applications like autonomous assistants, simulation control, or research assistants that need to remember events over weeks can benefit from the temporal and causal structuring.
• Plug‑and‑play – Since MAGMA sits between the LLM and the external datastore, existing services (OpenAI, Anthropic, etc.) can adopt it with minimal changes to the generation pipeline.

Limitations & Future Work

• Graph construction overhead – Building and maintaining causal and entity graphs requires additional annotation or a reliable classifier, which may be noisy in low‑resource domains.
• Policy learning data – The traversal policy is trained on synthetic or benchmark queries; transferring to highly specialized industry vocabularies may need further fine‑tuning.
• Scalability to billions of nodes – While the current experiments handle up to a few hundred thousand memories, scaling the multi‑graph structure to truly massive corpora remains an open challenge.

Future directions include exploring hierarchical graph abstractions, integrating retrieval‑augmented fine‑tuning of the LLM itself, and extending MAGMA to multimodal memories (e.g., images, code snippets).

Authors

• Dongming Jiang
• Yi Li
• Guanpeng Li
• Bingzhe Li

Paper Information

• arXiv ID: 2601.03236v1
• Categories: cs.AI
• Published: January 6, 2026
• PDF: Download PDF