AgentDB 3.8 adds ruv-fann as a runtime dependency. Its ruv-fann-wasm
dependency supplies the actual portable core and SIMD binaries. No native build
or optional accelerator is required for dense inference or training.
import { createDatabase } from 'agentdb/db-fallback';
import { FannModels } from 'agentdb/controllers/FannModels';
const db = await createDatabase('models.db');
const models = new FannModels(db);
await models.create('linear', {
schema: 'ruv-fann.dense-layers/1',
layers: [{ inputs: 1, outputs: 1, weights: [0], bias: [0],
activation: 'Linear', steepness: 1 }],
});
await models.train('linear', {
inputs: new Float64Array([0, 1, 2]),
targets: new Float64Array([1, 3, 5]), rows: 3,
epochs: 300, learningRate: 0.05, optimizer: 'sgd',
});
console.log(await models.predict('linear', new Float64Array([1, 2]), 2));
models.dispose();
db.close();Checkpoints persist weights, biases, dimensions and revisions. Optimizer moments are not persisted; each training call starts a fresh optimizer. Training uses MSE and supports Linear, Sigmoid, SigmoidSymmetric, Tanh and ReLU activations. It commits only a complete finite checkpoint. A concurrent update rejects the stale training commit. Model IDs cannot overwrite an existing checkpoint.
Inference reuses prepared WASM models and batch workspaces in a bounded LRU.
Every call checks the persisted version, including delete/recreate cycles, before
reusing cached state. Call dispose() before closing the database to free cached
WASM allocations. The default cache holds at most 16 models and 32 MiB.
MCP tools use flat row-major numeric arrays:
| Tool | Minimum profile | Purpose |
|---|---|---|
agentdb_neural_get |
readonly | Read checkpoint and revision |
agentdb_neural_list |
readonly | List metadata without weights |
agentdb_neural_predict |
readonly | Batched inference |
agentdb_neural_create |
learning | Create a checkpoint |
agentdb_neural_train |
learning | Train and commit a revision |
agentdb_neural_delete |
admin | Delete a checkpoint |
Explicit allow/deny settings apply as for other MCP tools. Unknown arguments, non-finite numbers and invalid dimensions are rejected. Requests are bounded to 128 rows, 1,000 epochs and 50 million parameter × row × epoch operations. Model JSON is limited to 8 MiB. These limits constrain synchronous WASM work; training is intended for small dense models rather than foundation-model fine tuning.