@mounaji_npm/learning-core

The neural weighting engine as a reusable, root-parameterized library, extracted from mcp-mounaji-rag. It learns which intents, collections and skills matter — from search frequency, feedback, recency decay and co-occurrence — and turns those weights into boost multipliers via a sigmoid.

Zero runtime dependencies (Node built-ins only).

What changed vs. the MCP's internal copy

• Root is a constructor argument (not process.env.YOAGENT_ROOT), so many agents can each own an isolated engine — e.g. an @mounaji_npm/agentfs learning/ layer.
• onEvent is injected instead of a hard-wired dispatchEvent.
• Adds exportState / importState for durability (Supabase mirror) and portability (Tier S export).

File names and formats are unchanged, so it reads/writes the same intent_weights.json, collection_weights.json, etc. the MCP already uses.

Usage

import { createLearningEngine } from '@mounaji_npm/learning-core';

const engine = createLearningEngine({
  root: '/data/agents/yoagent-abc/learning',
  onEvent: (type, payload) => bus.emit(`learning.${type}`, payload), // optional
  scopeContext: { agentId: 'abc' },                                  // optional
  config: { learningRate: 0.15 },                                    // optional overrides
});

engine.processSearchEvent({
  query: 'how does billing work',
  intent: 'find_convention',
  collections: ['mcp_conventions'],
  skillsHit: ['billing-service'],
  topScore: 0.82,
  resultCount: 5,
  feedback: 'positive', // optional: 'positive' | 'negative' | null
});

engine.getIntentBoost('find_convention', 1.2); // learned + base, clamped
engine.getWeightSnapshot();                     // all weights + boosts for a UI

Compose with agentfs

import { createAgentFs } from '@mounaji_npm/agentfs';
import { createLearningEngine } from '@mounaji_npm/learning-core';

const fs = createAgentFs({ root, agentId });
const engine = createLearningEngine({ root: fs.layerPath('learning') });
// engine state now lives inside the agent's portable bundle.

Durability

const state = engine.exportState();     // plain object → persist to Supabase
engine.importState(state);              // rehydrate a fresh/ephemeral engine

API

• createLearningEngine(opts) / new LearningEngine(opts) — { root, config?, onEvent?, scopeContext? }.
• processSearchEvent(params) — main entry point; updates all weights, records the pattern, tracks co-occurrence, emits a search event.
• updateIntentWeight / updateCollectionWeight / updateSkillWeight.
• getIntentBoost / getCollectionBoost / getSkillBoost / getLearnedKeywords.
• detectIntentWithLearning(query, rules) — base + learned + stem matching.
• learnKeywords / trackCoOccurrence / recordSearchPattern.
• getWeightSnapshot / getSearchPatternData.
• appendFeedback / readFeedbackLog.
• loadConfig / saveConfig, getDefaultLearningConfig().
• exportState() / importState(state).

Test

npm test   # node --test