ricable @ruvector/sona
Self-Optimizing Neural Architecture (SONA) providing runtime-adaptive learning with LoRA fine-tuning, EWC++ elastic weight consolidation, and ReasoningBank pattern storage for continual AI agent improvement.
Quick Reference
| Task | Code |
|---|---|
| Install | npx @ruvector/sona@latest |
| Import | import { SONA } from '@ruvector/sona'; |
| Create | const sona = new SONA({ learningRate: 0.01 }); |
| Adapt | await sona.adapt(input, feedback); |
| Predict | const pred = await sona.predict(newInput); |
| Save | await sona.save('./model'); |
Installation
Hub install (recommended): npx ruvector@latest includes this package.
Standalone: npx @ruvector/sona@latest
See Installation Guide for the full ecosystem.
Key API
SONA
The main self-optimizing neural architecture class.
import { SONA } from '@ruvector/sona';
const sona = new SONA({
learningRate: 0.01,
ewcLambda: 0.5,
loraRank: 8,
});
Constructor Options:
| Option | Type | Default | Description |
|---|---|---|---|
learningRate |
number |
0.01 |
Base learning rate |
ewcLambda |
number |
0.5 |
EWC++ regularization strength (0.0-1.0) |
loraRank |
number |
8 |
LoRA adapter rank |
loraAlpha |
number |
16 |
LoRA scaling factor |
dimensions |
number |
128 |
Internal representation dimensions |
reasoningBank |
boolean |
true |
Enable ReasoningBank pattern storage |
maxPatterns |
number |
10000 |
Maximum stored reasoning patterns |
adaptThreshold |
number |
0.01 |
Minimum loss change to trigger adaptation |
batchSize |
number |
32 |
Adaptation batch size |
Methods:
| Method | Returns | Description |
|---|---|---|
adapt(input, feedback) |
Promise<AdaptResult> |
Adapt to new input-feedback pair |
predict(input) |
Promise<Prediction> |
Generate prediction |
evaluate(inputs, labels) |
Promise<EvalResult> |
Evaluate on a dataset |
consolidate() |
Promise<void> |
Run EWC++ consolidation |
save(path) |
Promise<void> |
Save model and adapter state |
load(path) |
Promise<void> |
Load model and adapter state |
reset() |
void |
Reset adapter weights |
getPatterns(query, k?) |
Promise<Pattern[]> |
Search ReasoningBank |
getStats() |
SONAStats |
Get learning statistics |
LoRAAdapter
Low-Rank Adaptation layer for efficient fine-tuning.
import { LoRAAdapter } from '@ruvector/sona';
const adapter = new LoRAAdapter({
rank: 8,
alpha: 16,
targetModules: ['query', 'value'],
});
Constructor Options:
| Option | Type | Default | Description |
|---|---|---|---|
rank |
number |
8 |
Low-rank decomposition rank |
alpha |
number |
16 |
Scaling factor |
dropout |
number |
0.0 |
LoRA dropout rate |
targetModules |
string[] |
['query', 'value'] |
Modules to apply LoRA |
EWCPlusPlus
Elastic Weight Consolidation with online Fisher updates.
import { EWCPlusPlus } from '@ruvector/sona';
const ewc = new EWCPlusPlus({
lambda: 0.5,
gamma: 0.99,
});
Constructor Options:
| Option | Type | Default | Description |
|---|---|---|---|
lambda |
number |
0.5 |
Regularization strength |
gamma |
number |
0.99 |
Fisher information decay factor |
onlineMode |
boolean |
true |
Use online (incremental) updates |
ReasoningBank
Pattern storage with trajectory tracking for learned behaviors.
import { ReasoningBank } from '@ruvector/sona';
const bank = new ReasoningBank({ maxPatterns: 10000 });
await bank.store({ task: 'classify', input, output, reward: 0.95 });
const similar = await bank.search('classify emails', { k: 5, minReward: 0.8 });
Methods:
| Method | Returns | Description |
|---|---|---|
store(pattern) |
Promise<void> |
Store a reasoning pattern |
search(query, opts?) |
Promise<Pattern[]> |
Search similar patterns |
getStats() |
BankStats |
Get pattern statistics |
prune(minReward?) |
Promise<number> |
Remove low-reward patterns |
Common Patterns
Continual Learning Agent
import { SONA } from '@ruvector/sona';
const sona = new SONA({ learningRate: 0.01, ewcLambda: 0.5 });
for (const task of taskStream) {
const prediction = await sona.predict(task.input);
const reward = evaluateOutput(prediction, task.expected);
await sona.adapt(task.input, { reward, expected: task.expected });
// Periodically consolidate to prevent forgetting
if (taskCount % 100 === 0) {
await sona.consolidate();
}
}
Pattern-Based Retrieval
import { SONA } from '@ruvector/sona';
const sona = new SONA({ reasoningBank: true });
// Retrieve successful past patterns
const patterns = await sona.getPatterns('authentication flow', 5);
const bestApproach = patterns[0];
console.log(`Best approach: ${bestApproach.output} (reward: ${bestApproach.reward})`);
Checkpoint and Resume
import { SONA } from '@ruvector/sona';
const sona = new SONA({ learningRate: 0.01 });
// Train on task A
await sona.adapt(taskAInputs, taskAFeedback);
await sona.consolidate();
await sona.save('./checkpoint-a');
// Train on task B without forgetting A
await sona.adapt(taskBInputs, taskBFeedback);
await sona.save('./checkpoint-ab');
RAN DDD Context
Bounded Context: Learning
References
- API reference: See references/commands.md
- Full README
- npm