How It Works

Overview

MemLib is a memory API that sits between your AI application and your own infrastructure. You bring your own database and AI provider keys — MemLib orchestrates the memory pipeline.

What MemLib Does

1. Extracts facts from natural language using LLM inference
2. Embeds each fact into a vector for semantic search
3. Deduplicates against existing memories via cosine similarity
4. Resolves conflicts when new information contradicts old (e.g., "moved from NYC" → "lives in Berlin")
5. Retrieves memories by meaning with hybrid scoring
6. Synthesizes context paragraphs tailored to the current conversation

What You Provide

This is the Bring Your Own Database (BYOD) model. MemLib never stores your data — it connects to your database at request time and runs the memory pipeline.

---

Request Lifecycle

Every API call follows the same pattern:

Key details:

• Authentication — every request requires an API key (Authorization: Bearer sk_...)
• Project isolation — each API key maps to a project with its own database URL and AI provider config
• Stateless — the API connects to your database per-request and disconnects after

---

Components

MemLib is composed of several layers:

---

Data Model

Memories are stored in two tables in your PostgreSQL database:

memories table

Each row is an atomic fact with its vector embedding:

memory_events table

Every mutation is logged as an audit event:

This audit trail powers the diff feature — zero-LLM-cost changelog queries.

---

Supported Providers

MemLib supports multiple AI providers out of the box:

LLM Providers (for fact extraction + conflict resolution)

Embedding Providers (for vector search)

See the Providers guide for detailed setup instructions.

Next Steps

• Namespaces & Entities — understand multi-tenancy
• Smart Store — how the storage pipeline works
• Quickstart — get up and running in 5 minutes