New agentic memory framework uses 118K tokens per query. LangMem burns through 3.26M.
THE SO WHAT
Long-horizon agents are hitting a hard wall on context bloat—MRAgent’s 118K-token approach vs LangMem’s 3.26M shows that memory architecture is now a primary cost and latency driver. If you’re building agents, you need an owner for memory strategy the same way you have owners for retrieval and tools.
READ THE SOURCE
MORE FROM THE WIRE
Applied AITrump Admin releases Anthropic Mythos to be used by more than 100 US companies, agencies
Selective access to Mythos 5 for 100+ US companies and agencies turns top-tier models into a regulated strategic asset, not a generic cloud SKU. If you’re outside this circle, assume a capability gap and plan around model diversity, export controls, and jurisdictional risk in your stack.
Applied AIAnthropic accuses Alibaba of copying Claude by asking it millions of questions — and sets the stage for a new AI war
Turning model distillation into a legal and geopolitical fight raises the cost and risk of training via API-query mimicry. If your roadmap depends on student models trained on a competitor’s outputs, you need an IP and compliance review, not just a training plan.
Applied AIAnthropic’s Mythos 5 is back
Model access is now a negotiated asset with the US government, not just a product SKU. If you’re building on frontier models, treat policy risk like vendor risk—map which workloads break if access is throttled or gated to “select organizations.”
Applied AITrump Administration Allows Anthropic to Release Mythos to Select US Organizations
“Select US companies and government agencies” getting Mythos access creates a de facto tiered AI market—some actors get higher ceilings on capability and speed. If you’re outside that circle, assume you’re competing against counterparts with more powerful tools and adjust your defensibility story accordingly.