A research team has built a single memory system for long-term AI agents that outperforms the current approach of stitching together separate databases.
Current agent memory setups typically combine vector databases with graph databases — two different storage systems that rarely talk to each other cleanly. The cross-database reads slow things down and fragment information that should be linked. Mandol collapses this into one "memory-native" architecture built around structured semantic graphs. It layers raw memories at the bottom and abstract, aggregated summaries on top. Retrieval happens inside that single structure, with no LLM calls required during the lookup step — just deterministic routing and denoising.
The performance numbers are the headline: 5.4x faster retrieval and 4.8x faster insertion under ten concurrent queries per second, on consumer-grade hardware. On two standard long-term conversation benchmarks — LoCoMo and LongMemEval — Mandol posted the best overall accuracy among tested memory systems. That combination of speed and accuracy on accessible hardware matters because most agent memory research runs on data center assumptions.
Agent memory is quietly becoming the bottleneck in multi-session AI applications — every major lab is grappling with how to make models remember things across conversations without hallucinating or losing context. Mandol's bet is that unification beats optimization: rather than tuning each database separately, eliminate the seams entirely. Whether that holds at scale beyond the benchmark conditions is the next question nobody in the paper answers.