AI memory experiment recap: most clever designs fail

On October 3, 2026, a post on DEV Community took stock of a set of experiments on AI memory. The author recalls having argued in July that agents need governed memory: memory with its own lifecycle that decays, gets promoted, consolidates, and keeps provenance. The follow-up work set out to test that idea rather than restate it.

Across ten experiments, run on public benchmarks and on the author's own real conversations, the goal was to let Aura Memory decide for itself what is worth keeping. According to the post, each experiment had a protocol and a pass threshold written down in advance, so the outcome of every run was judged against a pre-committed bar rather than adjusted after the fact. The write-up describes itself as the honest version of the story.

The conclusion is that most of the ideas the author put forward earlier did not hold up, and the headline states plainly that most of the clever designs were dropped. Some of what he believed in July, in his own words, turned out to be wrong. The post does not disclose the individual protocols, the pass thresholds, or the numeric results of the ten experiments, and it does not list which designs survived and which were discarded, so those details cannot be verified from the available material.