Onboarding an engineer whose AI already knows the codebase's past
A new engineer joins your team. The laptop is working by lunch. The codebase takes longer, because code only shows what, never why. There is a retry wrapper around exactly one API call, a table duplicating data a join could fetch, a feature flag marked temporary since last spring. Each of those has a reason, and each reason lives in someone's head. So the first month becomes a slow interview. Post the question in a channel, wait for whoever remembers, pull a senior out of deep work, accept "I think there was an incident" as the answer.
Hand the new engineer an AI assistant, which is how AI dev teams onboard now, and only half of that goes away. The model reads the whole repo in minutes and is fluent in the what almost immediately. The why is as invisible to it as it was to its owner. Worse, models fill gaps with confident guesses, so by day two it proposes deleting the retry wrapper, since nothing in the code says what the wrapper is protecting.
That gap is what we built kgai for. It is an MIT-licensed Claude Code plugin, with a standalone kg CLI, that captures decisions at the write boundary. A skill and hooks record structural choices as the agent makes them, into an append-only, immutable decision log, and the graph projection is deterministic, so the same log yields the same graph on every machine.
For the new hire this rewrites day one. Their agent queries the graph before touching an area. Recall is lexical plus graph traversal, with no embeddings in the read path, and it returns only decisions in force, so the agent works from current rules rather than the whole archive. The archive stays reachable, though. Superseding a decision is an explicit link with a reason, dead ends stay queryable along with why they died, and "how did this get this way" has a dated answer, the same shift we described in git blame tells you what, never why.
Joining the team memory is a sync step, opt-in, through an S3 bucket you own. No daemon, no server, anywhere in the path. Every writer appends to a shard of their own, so parallel writes never cause a textual conflict, and the design holds far past any onboarding scenario. The largest archived run is 1,000,000 decisions across 30 writers' shards with decision lookups near 100 ms, though recall and free-text search are slower than that.
We watched this play out when a fresh hire challenged an old call in the new dev asked why we skipped Elasticsearch. A question that used to cost a meeting cost one query. That is the trade on offer. Weeks of asking around, or a store that answers before the first standup.
Try it on your repo. Two commands, MIT, nothing leaves your machine until you configure a bucket.
See how kgai compares to other tools on the comparison page, or browse the other notes.