claudeers.

🔓 unclaimed — this page was auto-generated from GitHub. Are you the creator?

Claim this page →
// RAG & Knowledge

llm-ledger

A verifiable, time-aware knowledge ledger your LLM keeps for you. An evolution of the LLM Wiki pattern, built on Claude Code.

Actively maintained
100/100
last commit 5 days ago
last release 13 days ago
releases 1
open issues 0
// install
git clone https://github.com/verbio-labs/llm-ledger

Verbio the Owl — LLM Ledger

LLM Ledger

by Verbio Labs
A verifiable, time-aware knowledge ledger your LLM keeps for you.

English · 한국어


A permanent markdown knowledge base that an LLM synthesizes and maintains for you — an evolution of Karpathy's "LLM Wiki" pattern, implemented on top of Claude Code.

But instead of accumulating prose pages like a wiki, it accumulates atomic "claims" — each with its sources, a confidence level, and a validity window in time. You do the sourcing and the asking; the LLM does every write, every verification, and every cross-reference.

No external skills or plugins. This folder alone runs in any Claude Code environment. Take it, fill it with your own data, and grow it as you go — a starter, not a final answer.


Why a "ledger," not a "wiki"

The original LLM Wiki's atomic unit was a synthesized prose page. Powerful — but three things were structurally impossible:

Wiki's limitationThe ledger's fix
No provenance — "who actually said this sentence?"Every claim requires sources (original quote + locator). No unsourced assertions.
No time-awareness — "since when was this true, and is it still?"Every claim carries valid_from / valid_until. As-of queries work.
Past lost on conflict — new sources silently overwrite old textNothing is overwritten. Facts are superseded or marked contested, never erased.

In one line: a wiki is mutable prose; a ledger is append-only and auditable. We don't overwrite facts — we supersede them and keep the past.

And it inherits the original's best idea — constant token cost. index.md is not a catalog; it's a router (MOC) that takes the intent of a question and decides which shard to open. Claims are indexed by topic, time, and confidence, so the per-question token budget stays flat as the ledger grows.

The difference at a glance (shipped as a live example)

"As of 2005, was Pluto a planet?"

  • A normal wiki: "Pluto is a dwarf planet" — overwritten by the 2006 fact, so it gives the wrong answer.
  • LLM Ledger: the old claim is preserved as superseded, so it answers "Yes — in 2005 it was a planet." — the correct as-of answer.

See it in the repo: 30-ledger/claims/pluto.md · 50-queries/2026-06-19-pluto-as-of-2005.md


Quick start

cd llm-ledger
claude

A usage guide pops up automatically on first run (a SessionStart hook).

  1. Add a source/ingest <URL · file · text> saves the original into 10-inbox/.
  2. Ledger it/compile extracts claims from the source, assigns provenance / confidence / validity, resolves conflicts (supersede or contested), refreshes topic views and indexes, then moves the original to 20-raw/.
  3. Ask/query Was Pluto a planet in 2005? retrieves and cites claims via two-stage routing.
  4. Check/audit scans for contradictions, low confidence, and temporal inconsistency.
  5. Trace/timeline pluto shows how the knowledge changed over time.

In inbox = not yet compiled. In raw = compiled.


Architecture (4 layers)

LayerFolderOwner
inbox (uncompiled queue)10-inbox/you drop sources here
raw (immutable originals)20-raw/filled by compile, LLM reads only
ledger (synthesized truth)30-ledger/written entirely by the LLM (append-only)
schema (operating spec)CLAUDE.md + 00-system/conventions.mdco-evolved by human + LLM
10-inbox/ → /ingest → /compile → 30-ledger/
                                  ├── claims/      (atomic facts — the source of truth)
                                  ├── topics/      (read-only views assembled from claims)
                                  ├── index.md     (router MOC)
                                  ├── indexes/     (by-topic / by-time / by-confidence)
                                  └── aliases.md   (canonicalization = routing keys)
                      /query    → Phase A routing → Phase B retrieve & cite → 50-queries/ file-back
                      /audit    → contradictions · low-confidence · orphans · temporal checks
                      /timeline → unfold claims chronologically; reconstruct as-of snapshots
                      originals move to 20-raw/ after processing (kept immutable)

A claim — the atomic unit

id: clm-2026-0002
statement: "Pluto is the ninth planet of the Solar System."
sources:
  - ref: 20-raw/pluto-tombaugh-1930.md
    quote: "it was announced as the ninth planet of the Solar System"
confidence: high
valid_from: 1930-02-18
valid_until: 2006-08-24      # ← temporal boundary
status: superseded            # ← preserved, not overwritten
superseded_by: clm-2026-0003

A topic page is a view assembled from claims like these, with a [^clm-id] footnote on every assertion. If a view is ever damaged, it can be regenerated from the claims at any time.


Commands

CommandWhat it does
/ingest {source}Save material into 10-inbox/ (no synthesis)
/compile [source]Extract claims + provenance/confidence/validity + conflict handling + indexes + move to raw
/query {question}Two-stage routed retrieval + cited synthesis + file-back
/audit [topic]Contradictions · low-confidence · orphans · index integrity · temporal consistency
/timeline {topic}Trace over time / reconstruct an as-of snapshot

Learn more


Lineage & credits

Inspired by fivetaku/llm-wiki and Karpathy's LLM Wiki gist, extending the concept toward "claims as the unit + provenance/confidence/time + conflicts preserved."

MIT License. Fork it freely.

🦉 Verbio Labs

// compatibility

Platformscli
Operating systems
AI compatibilityclaude
LicenseMIT
Pricingopen-source
LanguageShell

// faq

What is llm-ledger?

A verifiable, time-aware knowledge ledger your LLM keeps for you. An evolution of the LLM Wiki pattern, built on Claude Code.. It is open-source on GitHub.

Is llm-ledger free to use?

llm-ledger is open-source under the MIT license, so it is free to use.

What category does llm-ledger belong to?

llm-ledger is listed under devops in the Claudeers registry of Claude-compatible tools.

3 views
25 stars
unclaimed
updated 14 days ago

// embed badge

llm-ledger on Claudeers
[![Claudeers](https://claudeers.com/api/badge/llm-ledger.svg)](https://claudeers.com/llm-ledger)

// retro hit counter

llm-ledger hit counter
[![Hits](https://claudeers.com/api/counter/llm-ledger.svg)](https://claudeers.com/llm-ledger)

// reviews

// guestbook

0/500

// related in RAG & Knowledge

🔓

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 60-95% fewer tokens, same answers. Library, proxy, MCP server.

// ragheadroomlabs-ai/Python56,255Apache-2.0[ claude ]
🔓

A collection of notebooks/recipes showcasing some fun and effective ways of using Claude.

// raganthropics/Jupyter Notebook46,409MIT[ claude ]
🔓

✨ Light and Fast AI Assistant. Support: Web | iOS | MacOS | Android | Linux | Windows

// ragChatGPTNextWeb/TypeScript88,415MIT[ claude ]
🔓

Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant contex…

// ragthedotmack/JavaScript86,462Apache-2.0[ claude ]
Connectorlinks several projects together across the ecosystem · 10 connections
→ see how llm-ledger connects across the ecosystem