Adaptive Memory

Adaptive Memory is a pattern for creating AI that "learns" about the user over time without needing expensive RAG for every query.

The Strategy

Instead of treating memory as a "Search Engine" (searching for keywords), we treat it as a LIVING PROFILE.

1. Observe: Watch every interaction.
2. Reflect: Update the profile based on new information.
3. Inject: Put the entire profile into the system prompt.

This mimics Human Memory: you don't "search a database" to remember your best friend's name. You just know it because it's part of your mental context.

Implementation

We can build this using a simple Cencori Workflow.

1. The Trigger: chat.message

We use an async trigger so the user gets a fast response, while the "Memory Formation" happens in the background.

cencori.on('chat.message', async (event) => {
  await cencori.workflows.trigger('adaptive-memory-update', event);
});

2. The Observer Step

The workflow asks a "Reasoning Model" (like DeepSeek R1 or GPT-4o) to analyze the chat.

Prompt: "Extract deep observations about the user from this conversation. Focus on beliefs, goals, and recurring patterns. Output a JSON list of NewObservations."

3. The Reflector Step

This is the magic. We don't just append. We Evolve.

We load the current Profile and the new Observations, and ask the Reflector Agent to merge them.

Prompt: "Here is the Current Profile and New Observations. Update the Profile. Resolve conflicts. Delete outdated facts. Compress into a dense, narrative format."

4. Persist State

We save the new "Dense Profile" to cencori.memory.kv (Key-Value Store) under user_profile:{id}.

The Result

Next time the user chats:

const profile = await cencori.memory.kv.get(`user_profile:${userId}`);
 
const systemPrompt = `
You are a helpful AI.
User Context: ${profile}
`;

This achieves High-Performance Long-Context Recall because the "Profile" is a highly compressed, high-density representation of the user's entire history.