Overview

GUMem (General User Memory) is a user memory engine for Agents. It accepts conversation messages and behavior records, extracts reusable Memory, and recalls relevant context in later Agent turns.

Core capabilities

1. Behavior and conversation memory analysis: Cover user conversations, tool calls, business events, and behavior records so the Agent can extract user profiles, preferences, and intent.
2. Layered extraction for short-term and long-term memory: Use the ActionLogs / Messages -> Facts -> Summaries -> Topics memory hierarchy to keep factual evidence while increasing the information density of long-term memory.
3. Strong performance on long-context conversation memory: Achieve 92.9% Judge accuracy on the LoCoMo benchmark, making GUMem suitable for long conversations, cross-topic history, and time-related memory scenarios.
4. Developer friendly: Memory data can be managed by users, and custom WebHooks let developers adjust memory generation, filtering, and governance logic with full control over memory data.

Smallest integration path

1. Create a Session and bind it to your business user identifier.
2. Write Message input after conversation or behavior events happen.
3. Call Query Memory before the next Agent response.
4. Put formatted_context or structured Memory into the model context.
5. After the response is generated, write the new user message and assistant reply back to GUMem.

The order matters: recall, generate, then write back. This lets the Agent use Memory settled before the current turn instead of treating its current reply as historical fact.

Documentation entry points

• Quick Start completes the first integration with the Node SDK, Python SDK, or cURL.
• How GUMem Works explains Message, Facts, Summary, and Topic.
• Multimodal Content explains how text, image, and video content enters GUMem.
• Add Memory writes new Message input.
• Query Memory explains how to recall Topic, Summary, Facts, and recent Message context.
• User Case shows how GUMem fits into a typical Agent turn.