LLMory-palace

Memory Palace for Any LLM. Zero infrastructure. Works everywhere.

What is this?

A prompting technique that lets ANY LLM (web chat, CLI, local models) remember and navigate information using compressed keys - like the ancient Greek memory palace technique, but for AI.

How it works

Traditional approach (wasteful):

Context window filled with:
- Full conversation history (50K tokens)
- RAG results (20K tokens)  
- Instructions (5K tokens)
Space left: 125K tokens

LLMory-palace approach (efficient):

Context window:
- Compressed keys (500 tokens)
Space left: 199.5K tokens (99.75% FREE!)

Usage

# Install
npm install llmory-palace

# Create a compressed key
npx llmory create my-project planning "memory palace" "compressed keys"
# Output:
# @my-project|planning|2026-02-27
# → memory+palace+compressed+keys

# Generate index for any LLM
npx llmory index "@proj1|active|2026-02-27" "@proj2|planning|2026-02-26"

Works with ANY LLM

Web Chat (ChatGPT, Claude.ai, Gemini)

Paste this into any web chat:

# LLMORY PALACE INDEX
@my-project|active|2026-02-27 → memory+palace+compressed
@another-project|planning|2026-02-26 → keywords+here

What do you know about @my-project?

Claude Code CLI

import { createKey } from 'llmory-palace';

const key = createKey('my-feature', 'planning', ['hooks', 'capture', 'ollama']);
console.log(key);
// @my-feature|planning|2026-02-27
// → hooks+capture+ollama

Why it works

LLMs have vast knowledge in their parameters. We don't INJECT knowledge - we TRIGGER retrieval using compressed keys.

Like how "Eiffel Tower" triggers: Paris, iron, 1889, Gustave Eiffel, tower...

The compressed keys work the same way - they trigger associative retrieval from the model's training.

No infrastructure needed

• No vector database
• No embeddings
• No server
• No API keys
• No setup

Just prompts that work anywhere.

License

MIT