Writing

Introduces AEI as a unified architectural property — the guarantee that what an agent knows, does, and decides remains coherent, correctable, and aligned with user intent across time. Proposes prospective memory as the missing primitive connecting memory architecture to steerability.

A 6-cell framework — Knowing, Doing, and Deciding × session vs. persistent scope — mapping what current agentic systems have solved and where active research and engineering investment is deepening.

End-to-end quality in agentic tool use is a multiplicative function of three layers: contract quality, execution quality, and orchestration-time selection. Optimizing selection alone yields limited gains if the upstream layers remain unaddressed.

Four memory tiers — semantic, episodic, procedural, and prospective — mapped to enterprise agentic use cases. The session boundary stress test: if commitments don't survive a session end without the user restating them, you have a prospective memory gap.

A two-stage pipeline for agentic memory — async generation from task traces, then inference-time selection. Addresses the open problem of sourcing memory from non-human-bot interactions where signal density is far lower than in conversation transcripts.