Your teammate's AI solved this yesterday. Yours is solving it again today.
On Tuesday a dev's agent spends most of a session discovering that the payment provider double-delivers webhooks, so the handler has to be idempotent. It reads code, greps logs, pokes the sandbox, gets there. On Thursday a teammate's agent touches the same handler and starts the identical investigation from zero. Same codebase, same conclusion, paid for twice.
That's the quiet tax on AI dev teams. When the LLM does most of the work, most of what the team learns is learned by agents, and agents forget everything at the end of a session. Nobody notices, because every session looks productive on its own. The waste only shows up if you compare transcripts across machines, and nobody compares transcripts across machines.
What worked for us was writing the conclusion down at the moment it becomes a decision. kgai, the MIT-licensed tool we build, captures at the write boundary. A skill and hooks in the Claude Code plugin record structural decisions while the agent makes them, into an append-only, immutable decision log. What changed, why, what got rejected. Trivial edits record nothing, so the log stays signal.
The other half is recall. Before an agent edits an area, it reads the decisions that shaped that area. Recall returns only decisions in force, so the agent isn't wading through the whole history, and it's lexical plus graph traversal with no embeddings in the read path. Same store, same answer, on every machine.
So Thursday goes differently. The second agent reads "webhook handler stays idempotent, the provider double-delivers" before touching the file, and skips the investigation entirely. When the provider later fixes delivery and the team changes course, nothing gets edited. A new decision supersedes the old one through an explicit link with a reason, and the old one stays queryable.
Getting the log across machines is opt-in and runs through an S3 bucket you own. No server, no daemon, and per-writer shards mean parallel writes never cause a textual conflict. The mechanics are in how AI dev teams share what they learn. The same log is also what saves you when a person leaves rather than a session ending, which is when a senior leaves.
Count how many times your agents rediscovered the same fact this month. Then try the log on one repo, install is two commands.
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.