Ask HN: Is operational memory a missing layer in AI agent architecture?

Main idea: agents may be missing a reusable operational memory layer for things they learn by actually doing tasks over time — distinct from user memory, retrieval/RAG, and fine-tuning.

Examples include:

- tool quirks discovered during execution

- workflow patterns that repeatedly work

- environment-specific process knowledge

- failure modes that are expensive to rediscover

I’m calling the pattern “Agent Experience Cache” for now.

I’m mainly trying to pressure-test:

- whether this is truly a distinct category

- where it overlaps with episodic memory / trajectory storage / tool-use traces

- whether the failure modes and invalidation risks are framed correctly

Draft here:

https://docs.google.com/document/d/126s0iMOG2dVKiPb6x1khogldZy3RkGYokkK16O0EmYw/edit?usp=sharing