A knowledge base that writes, curates and improves itself

Every AI coworker forgets. Close the chat window, and the context, the decisions, the hard-won workarounds — all of it evaporates. The next person to ask a similar question starts from zero, and so does the assistant.

Codumentor is built around a different assumption: an organization’s collective knowledge shouldn’t live inside a single conversation. It should accumulate, the way a senior engineer’s judgment accumulates over years on a team — except automatically, for everyone, from day one.

That accumulation happens in four stages.

Learn

Curator agents run in the background, reading conversations, documents and git history as they happen. They’re not looking for anything specific — they’re extracting the same things a sharp new hire would pick up by osmosis over their first six months: naming conventions, prior decisions and the reasoning behind them, workarounds for that one flaky service nobody’s had time to fix properly.

Curate

Nothing gets written to memory as raw transcript. Each extracted insight becomes its own Markdown file with YAML metadata — deduplicated against what’s already known, dated, and committed to a dedicated git repository. This is the part that makes the memory trustworthy rather than just large: a fact you already knew doesn’t get restated a hundred times, and every entry has a paper trail back to where it came from.

Retrieve

A vector database indexes that memory and surfaces the right entries as Rapid Direction Cues — short, relevant context injected before the agent answers, not a firehose of everything it’s ever learned. The goal is a coworker that remembers the one thing that matters for the task in front of it, not one that recites its entire history at you.

Correct

Memory that’s never challenged eventually goes stale, so the loop doesn’t end at retrieval. Your team can vote on, flag, and review individual memory entries. Seniority weights the ranking, so knowledge from people who’ve been right before rises to the top, and entries nobody’s confirmed in a long time quietly fall away.

The result is an organizational memory that behaves less like a wiki someone has to maintain, and more like a colleague’s instinct — built from everything the team has actually done, corrected by the team as it goes, and available to every conversation from the very first message.