Mastra Memory Tops BEAM at 100K, 1M

This is a meaningful scaling result, but not universal proof of SOTA: Mastra wins at shorter long-context ranges while trailing at 10M tokens.

• –Observational Memory uses background Observer and Reflector agents to compress conversations into a stable, cacheable memory log.
• –BEAM’s long-context framing is more relevant to production agents than short-session recall alone.
• –Stable context can improve prompt-cache hit rates and make latency and cost more predictable.
• –The comparison uses published scores, so developers should reproduce results on their own workloads before choosing a memory system.
• –Mastra’s implementation is open source and integrated into its TypeScript agent framework.