GUMem Concepts
Concept Model
GUMem uses four public layers:
| Layer | Description |
|---|---|
Topic | The high-level organization layer and first recall entry point. |
Summary | Long-term memory supported by Facts and suitable for later Agent tasks. |
Facts | Traceable facts, preferences, plans, constraints, or events extracted from Message input. |
Message | Raw input, such as conversation messages, behavior descriptions, or business event text. |
Writes usually flow from Message to Facts, Summary, and Topic. Queries usually start at Topic, recall relevant Summary, and add supporting Facts and raw Message when needed.
Concept Pages
How GUMem Works
Read How GUMem Works to understand GUMem's characteristics, write path, recall path, and the relationship between Topic, Summary, Facts, and Message.
Multimodal Content
Read Multimodal Content to understand how text, images, and video should enter GUMem. The current Memory pipeline processes text-shaped Message input; images and video should be converted upstream into descriptions, metadata, transcripts, or external references.
Performance
Read Performance to understand benchmark reporting for LoCoMo, LongMemEval, Mem0 comparisons, and read/write performance.
Task Pages
- Add Memory: Write new Message input and trigger Facts, Summary, and Topic processing.
- Query Memory: Recall relevant Topic, Summary, Facts, and recent Message context.
- Update Memory: Update memory through correction Message input and governance flows.
- Delete Memory: Archive, delete, or stop recalling memory by provenance and lifecycle.
- WebHooks: Clean, audit, or sync data at Facts, Summary, and Topic processing points.
Terminology Boundary
GUMem docs do not require you to operate on internal data structures directly. Public docs use only Topic, Summary, Facts, and Message so callers can reason about source, purpose, and governance boundaries.
If you only want the first integration path, start with Quick Start.