⚙️ How It Works

LLM Memory Notes (LLMMN) provides persistent semantic memory for your coding agents (e.g., Claude Code, Codex). Instead of re-scanning the entire project each time, agents can save structured notes and retrieve them later when needed.

Conceptual Flow

[Coding Agent] ⇆ [LLM Memory MCP Server] ⇆ [Stored Notes]

The agent scans the codebase or project context and stores concise notes.
These notes are tagged with the project ID (root folder name) for context.
Later, the agent queries the memory to recall relevant details instantly.

Example Use Cases

Onboarding: Save an overview of modules and folders. Later ask “Where is the Maps integration?” and get direct references.
Refactoring: After changes, refresh notes so the agent recalls new file structure.
Integration Tracking: Record how external APIs (e.g., Google Maps, Stripe) are used.
Error Context: Persist issues or workarounds so future debugging is faster.

Example Note (JSON)

This is an example JSON payload to create a note linked to a memory.
The note is linked using the memory identifier (root project folder name) which maps to memories.identifier.

{
  "title": "Google Maps Integration",
  "memory_identifier": "llm-memory-notes",
  "content": "## Summary\nUses Google Maps JS SDK in `app/javascript/services/maps_service.js` to render maps on the Locations page.\n\n## Details\n- **Paths:** app/javascript/services/maps_service.js\n- **Modules/Classes:** MapsService\n- **Dependencies:** google-maps-sdk (JS)\n- **Related Features:** /locations page, marker rendering\n- **Tags:** maps, frontend, integration"
}

The server resolves memory_identifier to memory_id and stores the text in notes.content. AI-powered embeddings are derived from title and content and used for vector similarity search, not naive string matching (e.g. SQL LIKE).

Example Retrieval

Later, the agent can query notes using keywords. The server performs a vector similarity search to find the most relevant notes.
In practice, retrieval is performed via an MCP resource query, for example:

readMcpResource("llm-memory-notes://search/notes?query=docs%20page%20controller%20route&memory_identifier=llm-memory-notes&limit=5")

This returns a JSON payload with matching notes, including their id, title, content, memory_identifier, and similarity scores.

Vector similarity search finds results based on semantic meaning rather than exact word matches. Embeddings convert text into numerical vectors in a high-dimensional space, and the search identifies the closest vectors by meaning, enabling more accurate and context-aware retrieval. Learn more

Why It Matters

Without memory: agents re-scan the same files repeatedly, wasting tokens and time.
With memory: agents reuse persistent notes, respond faster, and provide more consistent answers.