Tag
A team reflects on six common structural failure points in AI builds: context, identity, decision memory, attention, write-back, governance, and economics, and offers a diagnostic tool based on their experiences.
The author describes inverting the textbook agent memory design from retrieval-on-demand to injection-first to avoid latency and confident empty-context errors, detailing the architecture and a dangerous self-poisoning failure mode with write-back.