Coding agents ignore instructions 90% of the time; shared memory helps

On October 3, 2026, a DEV Community post described one author's hands-on test of how often AI coding tools forget the instructions they are given. In the scenario he lays out, he tells Claude Code to deploy only from the release branch. An hour later he opens Cursor, asks it for a deploy script, and the script goes to main instead. The next day he starts a fresh Claude Code session and has to explain the same constraint all over again. As the post puts it, every AI tool he uses starts from zero, and each keeps its own scraps of memory, if it keeps any at all. The result is that the user repeats himself, and when he forgets to repeat himself, the instruction quietly falls apart.

The post's headline frames the problem as AI tools ignoring your instructions 90% of the time, with shared memory fixing most of that gap. The excerpt and summary, however, only show the concrete episodes above; they do not give the sample size or the method behind the 90% figure. The author's proposed fix is a shared memory layer that would hold these constraints in one place and apply them across tools, rather than leaving each tool to keep its own isolated notes.

As of this report, the matter is still at the stage of one author's test and proposal. No vendor response appears in the material, and there is no indication of whether any of the tools involved plans to support memory shared across different products.