@remnic/belief-ledger

A Remnic library for a personal belief and prediction ledger. It stores claims, predictions, and opinions as structured Remnic facts, retrieves prior claims that may conflict, asks a host LLM to classify the relationship, and tracks prediction calibration over time.

This package is intentionally host-neutral:

• It depends on @remnic/core public APIs.
• It does not import OpenClaw, Hermes, or direct OpenAI clients.
• Host adapters provide LLM access through LedgerLlmAdapter, or wrap Remnic's existing FallbackLlmClient with createFallbackLlmLedgerAdapter.

Basic Usage

import { StorageManager } from "@remnic/core";
import {
  BeliefLedger,
  RemnicLedgerStore,
  createFallbackLlmLedgerAdapter,
} from "@remnic/belief-ledger";

const storage = new StorageManager("/path/to/remnic-memory");
const store = new RemnicLedgerStore(storage);
const llm = createFallbackLlmLedgerAdapter(fallbackLlmClient);

const ledger = new BeliefLedger({ store, llm });

const result = await ledger.capture({
  text: "I think local-first memory tools will beat cloud-only memory tools by 2027.",
});

if (result.challenge) {
  console.log(result.challenge.question);
}

What Gets Stored

Claims are written as Remnic fact memories with:

• tags: belief-ledger, claim kind, status, and optional domain.
• entityRef: the first extracted entity when one exists.
• structuredAttributes: stance, confidence, domain, entities, deadline, evidence links, status, prediction outcome, and Brier score.

That confirms issue #869's upstream requirements through the public core surface: custom metadata, entity-scoped retrieval inputs, and supersession via StorageManager.supersedeMemory.

Host LLM Access

The package asks for a LedgerLlmAdapter rather than an API key. OpenClaw, Hermes, CLI, or app adapters can route the calls through their normal Remnic LLM path. For OpenClaw-backed Remnic runtimes, pass the existing gateway-backed FallbackLlmClient to createFallbackLlmLedgerAdapter.

Core Flows

• capture: extract a structured claim, persist it, retrieve prior claims, judge contradictions, and draft a Socratic challenge when needed.
• crossExamine: run retrieval and LLM judging against an existing claim.
• supersede, split, resolve, snooze, ignore: typed ledger actions suitable for host tool/function surfaces.
• scoreDuePredictions: find predictions past deadline, grade them from a verdict source or return a user-verdict prompt.
• reflect: compute calibration summaries, Brier score, flipped claims, and dormant topics.