MAGMA: Multi-Graph Agentic Memory for Long-Horizon AI Agents

Dongming Jiang, Yi Li, Guanpeng Li, Bingzhe Li · UT Dallas & Univ. of Florida · 2026
arXiv 2601.03236 Interactive page

One event, four graphsIllustrative example

Each event is stored once as a node, then linked four ways. Toggle edge types; hover a node to isolate its links.
Hover a node.

Write path: two speeds

Fast path never blocks on an LLM; the slow path infers causal and entity edges in the background.

Walk one query: “What instruments does Melanie play?”Illustrative graph

Intent → edge steeringIllustrative levels

Hand-set weight table, not learned. Only one range is stated: causal edges get weights ~3–5 for WHY. Other cells show qualitative boost only.
Hover a cell.

Reported results

Where the 0.700 comes from

LoCoMo has 1,986 questions; 446 are adversarial (unanswerable).

Cited but never runCoverage matrix

Hover a cell.

Missing controls

References