Memory

System Ready

Karpathy WikiKarpathy WikiKarpathy Wiki

Active Memory OS: A layered symbiosis integrating onde with the Karpathy Wiki pattern. Human-readable markdown meets agent-precise typed graphs and TOON density.

Terminal Inject

❯onde init --template karpathy-wiki

View Registry Source

Global Installs 21.4k

Nodes / Edges 4 / 3

Architecture Manifesto

Integrating onde with the Karpathy LLM Wiki pattern isn't a replacement—it's a layered symbiosis. onde becomes the deterministic, typed, journal-backed kernel, while the Karpathy skills become the human-facing, semantic, plugin-driven UI layer.

Here’s exactly how onde would accommodate, extend, and operationalize the Karpathy wiki skill pattern in production.

🧱 1. Architectural Mapping: Karpathy → `onde`

Karpathy Component	`onde` Equivalent / Adapter	Role in Hybrid System
`SCHEMA.md`	`schema.md` + `CLAUDE.md`	`schema.md` enforces typed graph rules; `CLAUDE.md` holds human-facing conventions
`wiki/pages/*.md`	Materialized views from CRDT journal	Human-readable cache; auto-rebuilt via `onde materialize`
`wiki/index.md`	TOON-dense catalog + graph index	Agent reads first; routes queries to `onde` or LLM context
`wiki/log.md`	`onde journal --tail` + Git commit history	Append-only event log; versioned via HLC
`skills/*.md`	`onde` CLI wrappers + skill adapters	Translate natural language triggers → deterministic `onde` pipes
Raw `raw/` sources	`onde ingest <file	url>` → journal append

Core Principle: onde owns the Single Source of Truth (journal). Karpathy owns the human-facing projection (markdown). Skills become adapters between the two.

🔌 2. Skill-by-Skill Accommodation

🟦 `wiki-init` → `onde init -t wiki`

onde init --template karpathy-wiki

What it does:

Creates raw/, wiki/pages/, wiki/assets/
Generates a default schema.md with Karpathy conventions:

## Node Types
### concept
  title: string @required
  type: enum(concept, entity, summary, comparison)
  sources: edge(source) @many
  related: edge(concept) @many
  confidence: enum(high, medium, low) @default(medium)

Generates CLAUDE.md with Karpathy workflow prompts
Initializes CRDT journal (journal/) + Git repo
Sets up skill adapters in .claude/skills/

🟦 `wiki-ingest` → `onde ingest + materialize`

Karpathy: Agent reads source → discusses → writes wiki/pages/ → updates index.md → appends log.md

onde Accommodation:

onde ingest raw/papers/transformer.pdf --extract --discuss

Parse: onde extracts claims, entities, citations → writes to journal as typed events
Discuss: Pipe to LLM for human validation: onde journal --last | ollama run llama3 "Summarize & flag contradictions"
Materialize: onde materialize --target wiki/pages/ generates flat markdown from journal
Backlink: onde lint --backlinks auto-adds [[links]] to index.md and related pages
Log: onde journal append --type ingest --source transformer.pdf

Key Gain: Ingest becomes idempotent & replayable. Delete wiki/? onde materialize rebuilds it perfectly.

🟦 `wiki-query` → Hybrid Query Router

Karpathy relies on LLM context for all queries. onde adds deterministic routing:

/wiki-query "What routing strategies are used in MoE models?"

onde Router Logic:

Check index.md (TOON format) for keywords
If query contains logical operators (blocks, depends, status, type) → route to onde find
If query is semantic/synthetic → route to LLM context
Return unified answer with citations

Examples:

# Logical → onde handles directly
onde find concept --related "mixture-of-experts" --format toon | grep routing

# Semantic → onde exports context → LLM reasons
onde export concept:mixture-of-experts | claude "Compare routing strategies"

Token Savings: onde filters 90% of irrelevant nodes before LLM sees anything. Context window used only for synthesis.

Vectors fail at hard reasoning. When an agent uses this topology, relations become deterministic traversing rules rather than fuzzy distance metrics. Ensure onde agent --sync is running when modifying nodes manually to keep the CRDT journal perfectly aligned.

Optimal Use Cases

❯Building an active memory OS
❯Merging human-editable Markdown with rigid AI topologies
❯Requiring TOON-dense indexes to save LLM context window

OS Folder Mapping -> Graph Representation

Zero DB Conf

Concept: Folders = Types, Markdown = Nodes

[agent-os] ~ /karpathy-world
├── schema.md
├── concepts/
│   ├── concept-01.md
│   └── concept-02.md
│
├── sources/
│   ├── source-01.md
│   └── source-02.md
│
├── summarys/
│   ├── summary-01.md
│   └── summary-02.md
│
├── comparisons/
│   ├── comparison-01.md
│   └── comparison-02.md

Inference Engine Maps:

/concepts/

@type(concept)

/sources/

@type(source)

/summarys/

@type(summary)

...and 1 more sub-types

onde_viewer

schema.md

12345678910111213

## Node Types

### concept

id: string @id @default(uuid())

title: string @required

type: enum(concept, entity, summary, comparison)

sources: edge(source) @many

related: edge(concept) @many

confidence: enum(high, medium, low) @default(medium)

### source

id: string @id

url: string

Entity_Example.md

## Concept: Mixture of Experts

type:: concept

confidence:: high

The routing layer relies on sparse activation.

edges:

sources: [[source-moe-paper]]

related: [[concept-attention]]

Typed Graph Topology

N0

concept

N1

source

N2

summary

N3

comparison

Kernel Parameters

Compatibility

onde Core v1.4+

Language Bindings

Bash, TS, Python

Target System Ready

Similar TopologiesMemory

Palace

Memory

06

Chat Memory:Passive diary & structured memory spanning temporal events. Ideal for conversational agents that need to recall past interactions accurately.

Topology

wing▶room▶drawer

Installs 14.2k

onde init -t mempalace

KARPATHY