Back to Projects

Lyla Memory MCP

Production

20+ tools. Daily use. Real persistence.

A production MCP server with 20+ tools — smart search, entity graphs, session persistence, self-editing context, temporal decay. Used daily in Claude Code and Claude Desktop sessions. The deployed implementation of the Recall AI architecture.

Architecture diagram coming soon

The Problem

Claude Code sessions start fresh every time. The context window fills and compresses. Projects with months of history — custody cases, business planning, technical architecture — need a persistent layer that doesn't forget.

Anthropic's MCP protocol provides the integration standard. I needed to build the actual memory system that runs on top of it.

The Solution

20+ MCP Tools

smart_search, search_entities, get_entity_graph, save_session_memory, save_topic, update_self_context, get_recent_insights, and more. Each tool maps to a specific memory operation.

Self-Editing Context

The AI can update its own context.json — correcting facts, adding discoveries, tracking state changes. Auditable updates with reasons logged for every change.

Daily Sync Pipeline

6pm cron job runs smart_ingest.js — processing voice journals, session files, and documents into vectors, entities, and relationships. Automatic memory consolidation.

Cross-Client Compatibility

Works in Claude Code CLI, Claude Desktop, Cursor, and any MCP-compatible client. Same memory, any interface. Tested daily across multiple workflows.

Tech Stack

TypeScript(Language)
Node.js(Runtime)
MCP Protocol(Protocol)
Supabase(Backend)
PostgreSQL(Database)
pgvector(Vector DB)
OpenAI Embeddings(AI)
Cron Jobs(Automation)

Key Decisions

Production, Not Prototype

This server runs every day. It powers custody case research, business planning, technical architecture work, and personal journaling. 324+ entities, hundreds of session insights, months of accumulated context.

Surprise-Based Storage

Inspired by Google Titans research. Only store what breaks patterns — if a memory is predictable from existing context, skip it. This keeps the memory system lean and meaningful instead of drowning in repetition.

Architecture from Recall AI

Recall AI is the research project — the 6-layer design, the temporal decay curves, the consolidation algorithms. This MCP server is the deployment — the TypeScript implementation that runs in production with real data.