LLM Wiki / OKF Research

Who made it and when

OKF v0.1 was published on June 12–13, 2026 by Google Cloud (Data Analytics team). Authors named in the blog post are Sam McVeety and Amir Hormati. The "Google" attribution in the Karpathy gist thread is accurate; this is not a community-driven or independent spec — it came directly out of Google Cloud Platform.

The format was explicitly framed as a formalisation of the emergent "LLM wiki" pattern that Karpathy's gist (and Obsidian / CLAUDE.md users) had independently converged on. Google's framing: "OKF standardises what practitioners already discovered."

Canonical locations

License: Apache 2.0 (confirmed by the Rust derivative work's NOTICE file; matches upstream).

Data model

bundle/
├── index.md          # reserved: directory listing; no frontmatter
├── log.md            # reserved: change history, date-grouped ISO 8601 entries
├── datasets/
│   ├── index.md
│   └── orders_db.md
└── tables/
    ├── orders.md
    └── customers.md

---
type: BigQuery Table
title: Orders
description: One row per completed customer order.
resource: https://console.cloud.google.com/bigquery?p=acme&d=sales&t=orders
tags: [sales, revenue]
timestamp: 2026-05-28T14:30:00Z
---

Versioning and evolution

• Current: v0.1 (Draft) — explicitly a starting point.
• Minor version bumps: backward-compatible additions.
• Major version bumps: breaking changes permitted.
• Contribution model: open PRs and issues on GitHub.

Tooling

Mapping trip2g → OKF (and back)

trip2g already uses a markdown-vault model that aligns closely. The concrete mapping:

Export: Add type: to every note's frontmatter (can default to the directory name or a configured mapping). Convert [[wikilinks]] to absolute bundle-relative Markdown links. Done — valid OKF bundle.

Import: Parse YAML frontmatter, treat type as a tag/filter dimension. Resolve Markdown links back to wikilinks lazily. Preserve all unknown frontmatter keys.

2.1 Concept identity / taxonomy drift (Synto — kytmanov)

What it is: Synto introduces a stable entity_id separate from display names. Surface forms change ("qubit" / "Qubit" / "qubits"), but the underlying concept identity does not. Curation commands (merge, split, rename, unmerge, keep) let an operator restructure the concept graph without breaking links or losing sources.

• merge LOSER WINNER [--absorb-edits]: moves sources and edges to winner, retires loser article, keeps loser labels as aliases.
• split NAME --sense SENSE SOURCE_PATH ...: partitions a concept's sources across senses; creates a disambiguation page at original name.
• keep SURFACE ENTITY: resolves homonym ambiguity, updates state DB.
• Supports --dry-run and synto undo.
• No vector DB required; works with local LLMs (Ollama / LM Studio).

Why it matters: Without stable IDs, rename + merge operations silently break cross-references. At scale (hundreds of concepts) homonym drift is the main cause of inconsistency.

Source: https://github.com/kytmanov/synto

2.2 Synthesis decay / hallucination-as-truth mitigation (Eidetic — LARIkoz; OpenClerk — yazanabuashour)

2.3 Contradiction detection without O(n²) (hmbseaotter)

What it is: Three-tier approach that keeps contradiction checks cheap at scale by exploiting the wikilink graph structure.

Tier 1 — Per-source detection at ingest (high frequency, ~8–15 pages):

• Compares incoming claim against only the pages touched by this ingest.
• Classifies as: soft / scope-mismatch / hard / none.
• Hard contradictions stop ingestion and block the commit.
• Machine-readable severity token in the file: "Contradiction severity: hard / Status: Unresolved — flagged for user review".

Tier 2 — Deterministic commit gate (zero LLM cost):

• Python os.walk + regex grep for "Status: Unresolved" — no context window needed.
• Runs before every commit; scales to any wiki size via cheap disk I/O.

Tier 3 — Periodic lint backstop (scoped to graph neighbourhood):

• Deterministic checks (orphan detection, missing-page detection, unreferenced images) via shell/grep — no LLM.
• LLM invoked only for reasoning-heavy cases: contradiction backstop, causal-chain gaps, thin-page judgment.
• Contradiction surface bounded by wikilink graph: only check nodes changed since last pass + their 1st- and 2nd-degree link neighbours. "Contradiction can only exist between claims about the same entity or relationship" — the graph hands you the neighbourhood for free.

The intentional gap: A contradiction between two old unchanged pages that have never landed in the same lint neighbourhood may slip through until a periodic full sweep. This is an explicit tradeoff: the per-source check and commit gate filter the real-world corruption cases; rare cold contradictions are caught by periodic sweeps.

Source: Karpathy gist comment thread — https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

2.4 Boundaries: canonical source / derived index / agent write authority (OpenClerk + distorx)

OpenClerk (see §2.2): canonical markdown → derived SQLite → agent write authority, with doc_id/chunk_id provenance.

distorx dual-index (production at ~7,500 notes):

• FTS5 trigram index for keyword/substring search: ~0.07 s latency.
• 1024-d vector embeddings (mxbai-embed-large via Ollama) for semantic search: ~1.1 s cold.
• Content-hash authority: embedding and re-index triggered by content hash change, not mtime. Prevents spurious re-embeds on filesystem touch events; prevents vector-space mixing when the embedding model is swapped (embedding-identity guard).
• MOC hubs (Map of Content): written once per area via graph analysis; member notes don't churn. Reduced orphan rate from 31% to 19%.

Source: Karpathy gist comment thread — distorx comment; https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

2.5 Concurrency-safe writes (Matryca-Plumber — MarcoPorcellato)

What it is: Shifts the storage model from flat Markdown text to a block-based outliner AST (Logseq format) with UUID per block. LLMs target block UUIDs rather than guessing where to inject text. Write safety via Optimistic Concurrency Control (OCC):

• Snapshot st_mtime before inference.
• Acquire page lock (page_rmw_lock + platform flock) only for the write.
• If the file was edited while the model was thinking, the commit aborts: "your typing always wins."
• Single shared mutation plane (graph_dispatch) — CLI commands, MCP tools, and background daemon all use the same write path, eliminating race conditions between operators.

Source: https://github.com/MarcoPorcellato/matryca-plumber; https://github.com/MarcoPorcellato/logseq-matryca-parser

2.6 Memory-as-external-MCP-service positioning (Smriti-MCP — deepak-bhardwaj-ps)

What it is: Memory externalised from the agent lifecycle as a standalone MCP server. Multiple agents (Claude, OpenAI, local models) share one wiki by pointing to the same --memory-root directory. Memory survives model changes and agent restarts because the storage is plain markdown with YAML frontmatter. Authorship tracked via optional source_agent metadata field.

12 MCP tools exposed: create_memory, get_memory, append_memory, update_memory, delete_memory, remember, consolidate_memory, search_memory, list_memories, archive_memory, build_memory_index, rebuild_memory.

Source: https://github.com/deepak-bhardwaj-ps/smriti-mcp

2.7 Typed relationship edges via seeded ontology (equationalapplications)

What it is: On top of OKF's untyped edges, three configurable modes for extracting typed relationships:

1. Strict (Seeded): Allowed edge types supplied upfront (e.g., ['client', 'employed_by']). LLM forced to stick to schema — guarantees predictable data structures for hard-coded dashboards. Maps to okf_type column + llm_wiki_edges table.
2. Emergent (Autogenerated): LLM freely invents relationship types, tracked in an ontology_manifest. Maximum expressivity, unpredictable schema.
3. Off (Disabled): Semantic fact extraction only, no relationship traversal.

A parseOkfBundle adapter is in development for consuming foreign OKF bundles and projecting them into this typed-edge model.

Source: Karpathy gist comment thread; https://github.com/equationalapplications/expo-llm-wiki