ai-secretary-kit automates AI memory upkeep after 4 months

The kit is a response to a problem that only appeared after months of daily use. The first version of the memory box held just three parts, but more than 1,400 commits later the assistant read so much material before speaking that sessions had grown slow. Adding new rules made things worse: the same type of mistake recurred a third time even after two rounds of writing extra rules, which is why the author converted repeated failures into checks instead.

A second pressure came from the system's own growth. The author estimates that files always read before a session began reached about 14,000 tokens, while three logs of feedback and failures totaled about 101,000 tokens. Notably, the rule to clean old records quarterly existed in writing, and its first deadline passed without any action, because clean-up was conditioned on the heavy task of confirming nothing had been lost. The fix was to set size limits, let a gate block saving when limits are exceeded, and move old records with a single command.

The kit ships with 14 distilled operating rules, drawn from 118 pieces of feedback and 113 failure records over four months. The author notes the kit was transcribed from an actual repository with personal content removed, and that the figures come from measurements there, with token counts as approximations. Not every part is automated: checks cover structure, such as broken references and oversized files, but not whether the content is correct, and stale information still gets marked and reviewed monthly by a person and the AI.