BanditDB

AI agents now pick which model to call, which offer to show or which customer to phone, thousands of times a day, with no record of what each choice cost or whether it worked. BanditDB stores every context, action and outcome, learns from them, and chooses the next action within the budget a team sets, so each decision is explainable and gets better.

A Postgres database answers what is true, a vector database answers what is similar, and nothing in the stack answers what to do. So when an agent chooses between an SMS, an AI call and a human call for a delinquent account, the choice is a black box: it never learns who paid, it cannot explain its confidence, and it has no idea that the team can only afford two thousand calls this month. For low-value decisions that is tolerable. For decisions that spend real money it is not.

BanditDB is an in-memory database built around one loop: a context, an action, a reward. Define the decision, ask for the best action for this context within the budget, report what happened, and the model updates in place, with the causal lift of each action measured against a baseline. It fits any repeated choice from a fixed set with a measurable outcome, from LLM routing and ad creatives to retention offers and collections, and a built-in agent prunes weak actions and proposes new ones. It is open source and self-hostable, with a pay-per-decision cloud in beta. The engine is a spinoff of Dynamic Pricing AI, where Simeon Lukov and Georgi Angelov ran bandit models in production for live pricing before turning the patchwork of tools into one database.