AllocateWork in. Weights out. Forever.
RecursiveDB is the recursive-learning engine underneath Allocate: it captures every task your agents complete, binds it to the outcome, and turns the result into training signal. Rented models stay static. Yours compounds.
Recursive learning is the loop where a model's completed work becomes its own training data. RecursiveDB is Allocate's implementation: every task an agent finishes is captured with the outcome that followed it, auto-labelled into a training-ready set, and folded into the next version of a model you own. Usage becomes accuracy; accuracy becomes your moat.
Loop
A task completes
An agent clears a dispute, files a document, answers a member. The full trajectory, prompt, tools, context, and output, is captured losslessly at the token level.
Learn moreThe outcome binds to it
What happened next is the label: the analyst agreed, the booking held, the ruling was upheld or reversed. Rewards are real outcomes appended to real work, never synthetic scores.
Learn moreSignal becomes a training set
Trajectories and outcomes are auto-labelled toward a training-ready dataset, alongside anything you push through the data tunnel: historical cases, documents, resolved records.
Learn moreTraining produces your next version
Fine-tuning runs against an open-weights base turn the dataset into a new model version, attributable to its data and reversible. The weights land inside your boundary, versioned.
Learn moreWhat you own.
Your data
Everything captured stays inside your isolation boundary and trains only your model. Exportable at any time; residency is the architecture, not a policy.
Your weights
Fine-tuned open-weight models belong to you contractually. If you leave, the weights and the dataset leave with you.
Never a public store
RecursiveDB surfaces are read-only views and export. The store itself is never a public query API, so the training corpus is not an attack surface.
Versioned to the token
Model, prompt, tools, and environment carry an immutable version on every trajectory, so any improvement is attributable and any regression reversible.
Common questions
What is recursive learning?
Recursive learning is the loop where a model's completed work becomes its own training data: tasks are captured, bound to real outcomes, and folded into the next fine-tune. Each version does the work that trains the version after it, so accuracy compounds with usage instead of staying flat.
What counts as training signal?
Recorded outcomes on real work: an approved claim, a confirmed booking, a human correction, a ruling upheld on review. Signal is appended, never edited, and bound to the exact trajectory that produced it.
Who owns the data and the trained model?
You do. The captured dataset and the fine-tuned weights live inside your isolation boundary, are contractually yours, and are exportable at any time. Your data never trains anyone else's model.
How is this different from fine-tuning alone?
Fine-tuning is one run over a dataset you assembled once. Recursive learning is the system that keeps assembling the dataset: capture and outcome-binding run on every task, so each training run starts from more and better-labelled data than the last, without a labeling project.
What exists today?
Capture, outcome binding, the data tunnel for historical uploads, and read-only views with export ship today. Training runs are scoped with our team per customer rather than a self-serve button; every run lands weights in your boundary under the same ownership terms.