Autonomous agent pipelines increasingly rely on knowledge graphs to maintain long-term memory across complex enterprise workflows. Yet most production stacks still dump extracted facts into flat triple stores, creating an un-auditable mess where no single component owns a statement and nobody can trace the evidence behind it. To tackle this organizational decay and hallucination creep in agent memory, a University of Maryland research team led by Pranav Bykampadi introduced MAGG, a multi-agent framework built to enforce Governed Knowledge Graphs.
Multi-Layer Architecture and Domain Ownership
The MAGG architecture splits the memory lifecycle across three coordinated layers: creation, governance, and application. At the creation layer, the system digests raw documents through a multi-stage extraction pipeline that uses blackboard-style deliberation to resolve ambiguous entities, relations, and candidate triples. Instead of forcing ML teams into rigid manual ontology engineering, a domain classifier induces entity and relation types on the fly, enabling the pipeline to operate in dynamic, open-world settings without predefined schemas.
Once candidate triples emerge, the governance layer routes them to specialized Domain Expert Agents. These are not human bottleneck reviewers, but dedicated LLM agents that inspect proposed facts against source evidence and local graph context before issuing an explicit admission decision.
Candidate triples are assigned to domain owners, reviewed against supporting evidence, admitted through governance decisions, and stored with audit metadata.
Because the resulting graph embeds explicit ownership boundaries, the application layer reuses this structure during downstream retrieval. When processing user queries, the system routes requests directly to domain-specific graph experts rather than relying on noisy, undifferentiated vector searches across an unpartitioned database.
Empirical Results on Extraction and Multi-Hop QA
Benchmark evaluations show tangible gains in both extraction fidelity and multi-hop reasoning over conventional baselines. On the SciERC benchmark, MAGG boosted strict triple F1 by 47% and mapped triple F1 by 51% compared to unmanaged flat insertion. A blinded evaluation of 120 triples confirmed that revised triples in MAGG matched source evidence in 100% of cases, substantially outperforming flat extraction in empirical support.
On the complex reasoning benchmark MuSiQue, MAGG outperformed Microsoft GraphRAG by 9.0 exact-match points and 11.2 token-F1 points. The multi-agent review layer effectively filters out hallucinations before they corrupt the shared memory substrate, ensuring that subsequent multi-hop steps build strictly on verified triples.
For enterprise engineering leaders, MAGG proves that baking governance directly into graph construction cures agent memory degradation without the prohibitive tax of manual schema maintenance. Replacing dumb triple stores with specialized LLM reviewers delivers verifiable provenance and sharper multi-hop retrieval. However, running multi-agent deliberations and verification layers inflates compute overhead during ingestion—a trade-off engineering leads must weigh carefully before plugging it into high-throughput streaming pipelines.