Skip to content
Core LLM Wiki

Introduction

Core LLM Wiki is a family of TypeScript packages that give AI agents persistent, structured, long-term memory, stored in SQLite. It sits between raw conversation logs and a curated knowledge base:

  • You log what happens (events) and feed it documents.
  • An LLM pass called the librarian turns those into durable facts and tasks.
  • At answer time you read() the most relevant memory and inject it into your prompt.

It was inspired by Andrej Karpathy’s LLM Wiki idea. “LLM wiki” has since become a general name for that pattern. Core LLM Wiki’s packages are published under the @equationalapplications npm scope, such as @equationalapplications/core-llm-wiki, @equationalapplications/react-llm-wiki, and @equationalapplications/expo-llm-wiki. Core LLM Wiki can import and export the Open Knowledge Format (OKF), so your knowledge isn’t locked in.

What you get

Section titled “What you get”
Capability What it means for you
Episodic → semantic memory Events consolidate into facts when maintenance runs, so you never hand-write a summarizer.
Hybrid retrieval Vector search when you provide embed(), keyword search otherwise (offline-safe), or a weighted blend.
GraphRAG Typed edges between facts, walked with one recursive SQL query. No graph database required.
Multi-entity Many independent “brains” in one file (per user, per agent, per tier), readable together with weights.
Provenance Every fact records where it came from. Document facts can’t be rewritten by the model.
Trust controls Draft review, evidence grounding, and lint reports, all opt-in.
Observability A typed, content-free onDiagnostic stream reports anything the engine drops.

The packages

Section titled “The packages”

The packages are published from one monorepo and versioned together.

Package Use it when…
@equationalapplications/expo-llm-wiki You’re building an Expo / React Native app.
@equationalapplications/react-llm-wiki You’re building a React web app.
@equationalapplications/core-llm-wiki You’re on Node.js, plain JS, or another framework. This is the engine all the others build on.
@equationalapplications/core-llm-tools You want Gemini function-calling schemas with scope-based authorization.
@equationalapplications/core-okf You need raw OKF parse/produce primitives.
@equationalapplications/schema-org-llm-wiki You want a ready-made schema.org ontology for your graph.
@equationalapplications/schema-software-org You want an ontology for a software organization’s agent.
@equationalapplications/prisma-outbox You need to sync memory to a Prisma-backed server database.
@equationalapplications/mcp-server You want to give coding agents (Claude Code, Cursor…) hosted SynapseTree memory.

Not sure which package you need? See Choose a package.

Used in production

Section titled “Used in production”

Core LLM Wiki powers Curated Thoughts, a desktop second brain whose integrations share one memory across your AI agents. It also powers Curated Journal, an offline mobile journal, and SynapseTree, a hosted cloud memory service. See Ecosystem.

Design principles

Section titled “Design principles”
  • Bring your own inference (BYOI). Core LLM Wiki never ships an LLM client or API key. You supply generateText and, optionally, embed.
  • Namespace safe. Every table is prefixed (llm_wiki_ by default), so it can live inside your existing app database.
  • No silent failures. Anything the engine drops (a bad chunk, a rejected fact, a dangling edge) can be reported to you through onDiagnostic.
  • Safe defaults. New behavior is opt-in. Upgrading a minor version doesn’t change what your app stores.