@cognivia/core

Cognivia's Learning Genome + Session Report engine. Zero dependencies. The same algorithm behind the site, the developer console, the CLI, the MCP server, and the hosted /v1 API, so every surface returns identical diagnostics.

import { buildReport, renderText } from "@cognivia/core";
const report = buildReport(attempts);   // attempts = [{subject,question,answer,correct,latencyMs,confidence}]
console.log(renderText(report));

@cognivia/core/client is a thin client for the hosted API (set COGNIVIA_API_KEY).

The stateful genome (GET /v1/learners/:id/memory-state)

With a live key, memoryStateRemote(id) returns each concept's full seven-variable Learning Genome and a plain-language diagnosis:

import { memoryStateRemote } from "@cognivia/core/client";
const ms = await memoryStateRemote("alice");
const c = ms.concepts[0];
c.genome.lambda.value;        // forgetting rate (Bayesian posterior mean)
c.genome.lambda.ci95;         // [lo, hi] 95% credible interval
c.genome.retrieval_strength.value;
c.diagnosis.headline;         // e.g. "Confident misconception"
c.diagnosis.action;           // what to do about it
ms.genome.lambda_weighted_mean;  // whole-learner signature

Per concept the genome carries lambda, retrieval_strength, latency_index, confidence_gap, consolidation_efficiency, fatigue_susceptibility, and pattern_dominance, each with a basis naming its science. Concepts with too little history return status: "insufficient_evidence" and no genome. The full schema is in docs/openapi-v1.yaml.

MIT · Manik Maurya