Core_Concepts

How it Works

Understand the primitives that power Memori's active state management.

1. Active Memory Layer

Unlike a traditional vector database where you push/pull vectors manually, Memori acts as a middleware. It sits between your application code and the LLM, intercepting requests to inject relevant context.

App Code──►Memori Layer──►LLM

2. Auto-Augmentation

This is the "magic" part. When you call register(client), we monkey-patch the client's completion methods/hooks to add a pre-processing and post-processing step.

Retrieval (Pre-flight)

The user's prompt is embedded and compared against the local vector index. Relevant memories are found.

Injection

Context is inserted into the system message, invisible to the end user but visible to the model.

Storage (Post-flight)

The resulting conversation turn is saved, embedded, and indexed for future reference.

3. Attribution & Scoping

For multi-tenant applications, you need to ensure User A essentially never sees User B's memories. Memori handles this via Attribution.

typescript

1// Global scope (default)
2const memori = new Memori();
3
4// User scope
5memori.attribution("user_123");
6await memori.addMemory("My name is John"); 
7
8// Switch user
9memori.attribution("user_456");
10const memories = await memori.search("What is my name?"); 
11// -> Returns [] (Empty, because John's memory is isolated)