dsh-cognitio

AI memory and self-discipline are, at the end of the day, a matter of model goodwill. Change sessions, swap presets, and the rules and lessons you have corrected again and again can simply be forgotten. dsh-cognitio turns the idea of remembering your rules and reminding you at the right moment from a gamble into a deterministic mechanism.

It works in three layers. The memory layer stores facts, rules, patterns, cases, and propositions in a tiered SQLite store with a full version chain, so every edit remains traceable. The trigger layer is a set of mechanical sentinels that fire at four deterministic points-turn start, tool call, error, and turn end-retrieving relevant rules into the context, completely decoupled from whichever model or preset is active. The evolution layer is an error-driven pipeline: mistakes are auto-archived as drafts, a daily unattended maintenance pass consolidates and distills them, and any newly proposed rule lands in an approval queue. Nothing takes effect until you approve it, and every rule carries a performance scorecard tracking how often it was useful versus how often it was stale. Repeatedly failing rules are automatically demoted into a rebuttal pool for review.

It is meant for people who collaborate with AI long-term, who find themselves correcting the same behaviors over and over, and who have accumulated a body of project rules they do not want to re-teach every conversation. This is not a chat bot-it is a self-evolving rules and memory system that keeps every byte of data on your local machine, fully auditable and fully under your control.