BanditDB is the decision layer for AI agents and adaptive applications.
It learns from context, actions, and outcomes to select better models, tools, prompts, offers, and workflows over time.
Static rules cannot adapt to changing users, tasks, costs, or outcomes. BanditDB turns repeated decisions into a measurable learning loop.
- # rules drift and exceptions multiply - if task == "coding": - model = "model_a" - elif customer == "enterprise": - model = "model_b" - elif cost_limit < 0.02: - model = "small_model" - else: - model = "model_c"
Every new condition is another branch you maintain by hand.
+ # learn from every measurable outcome + arm, iid = db.predict( + "model_router", + request_features + ) + + response = run(arm) + + db.reward(iid, quality_score(response))
One call. The policy updates itself from real outcomes.
Store facts and transactions
Retrieve related information
Learn from decisions and outcomes
your app / agents โ โโโ SQL โโโโโโโโโโโโโโบ postgres "what is true?" โโโ embed + ANN โโโโโโบ vector store "what is similar?" โโโ predict/reward โโโบ banditdb "what should I do?" swap any layer independently ยท plain HTTP ยท no coupling between stores
Good fit: a finite set of actions, repeated often, with a measurable reward. Bad fit: one-off decisions, unbounded action spaces, unmeasurable outcomes.
| domain | context | action | reward |
|---|---|---|---|
|
Insurance
|
claim complexity, prior history, policy tier | route to fast-track, adjuster, or SIU review | time to resolution |
|
Ad tech
|
device, time of day, audience segment | which creative variant to serve | click or conversion |
|
Logistics
|
order size, zone congestion, SKU mix | pick path or packing station | fulfillment time |
|
LLM routing
|
prompt embedding, task type | which model handles the request | task success, cost-adjusted |
|
Checkout
|
cart value, returning visitor, device | discount, free shipping, or gift | completed purchase |
|
Banking
|
credit utilization, income, repayment history | credit line offer to extend | repaid without default |
|
Healthcare
|
symptoms, vitals, treatment history | which protocol to recommend | patient outcome improved |
|
Edtech
|
skill level, past attempts, learning pace | which exercise or hint to serve next | concept mastered |
Two objects hold the learning state: a matrix tracking evidence and uncertainty, and a vector linking observations to rewards. Together, they calculate the best action fresh on every requestโnot retrieve it from a log.
Choosing an Algorithm โEvery decision keeps one matrix and one vector that update in place โ not a growing log โ so memory stays flat no matter how much traffic the DB has seen.
Observable over HTTP, durable through the Write-Ahead Log (WAL) and checkpoints, recoverable after a crash.
Prometheus metrics, per-campaign health diagnostics, and alerts if a campaign stops exploring properly.
A reward returns only once it is fsynced to the Write-Ahead Log โ an acknowledged write survives a crash a millisecond later. Checkpoints snapshot state, export to Parquet, and rotate the log automatically.
On restart, BanditDB loads the last checkpoint and replays everything written after it โ 100,000 events in about 0.6 seconds. If the process is killed outright, no acknowledged reward is lost.
The MCP server exposes campaigns as tools, so every agent in the fleet reads and writes one shared, persistent policy.
Two commands and BanditDB is a native tool in Claude, Cursor, or any MCP-compatible host โ no config file editing required. Not a chat memory, not a vector store: every recorded outcome sharpens one shared policy the whole fleet reads from.
Ordered simplest โ most advanced. Each ships runnable code and a reward-design discussion.
Temperature, light, or noise โ which adjustment works best for each person? A pure curl walkthrough, no SDK needed.
Discount, free shipping, or nothing โ learns which checkout offer closes each shopper without giving margin away.
Consult, intake form, refer, or decline โ learns which response maximises matter value for each enquiry profile, accounting for capacity and conflict risk.
Hold margin or liquidate? Learns from sell-through rate, holiday proximity, and competitor pricing โ context describes the market, not the user.
Learns which prompt strategy โ zero-shot, chain-of-thought, few-shot, structured โ produces the best response for each task type. Your evals run in production, not in a spreadsheet.
Routes patients toward the most effective treatment arm in real time as evidence accumulates โ no waiting months for interim analysis.
No sign-up. No cloud account. No configuration required.
binary โ linux, macos, windows
docker
Ask questions, share what you're building, get early updates. The BanditDB Discord is where the conversation happens.
Get help, share projects, talk to the team. Free and open to everyone.
Join DiscordQuestions about enterprise use, integrations, or just want to say hi โ book a 30-min call.