Mastra’s Observational Memory Tops BEAM

This is a strong result for memory architectures, but the 10M-token drop shows retrieval—not summarization—is the real frontier.

• –Observer and Reflector agents maintain chronological notes while a recall tool retrieves older evidence.
• –The system passes roughly 40K tokens to the agent even when conversations reach 10M tokens.
• –BEAM-10M performance falls to 66.4%, near the 68.0% frontier but within a wide error margin.
• –Major failures involve incomplete search, counting across sessions, and confusing assistant-generated advice with user statements.
• –Results are self-reported and depend partly on the answering model, making independent reproduction important.