Mastra Observational Memory: Dual-Agent Compression with Traffic Light Priority Encoding Achieves 95% on LongMemEval

Mastra Blog·high signal

Mastra's open-source observational memory uses two background agents — Observer and Reflector — to continuously compress conversation history into a dated log using red/yellow/green emoji priority markers that LLMs parse with unusual efficiency. This architecture enables 5–40x compression for tool-heavy agents and unlocks provider-side prompt caching (normally unavailable to RAG-based agents), cutting token costs by 4–10x. Benchmarks show 94.87% on LongMemEval with GPT-4o, surpassing the previous SOTA by 2.6 points.