What is a MemScene?
A MemScene represents a specific “context” or “theme” in the user’s life or work. It is a dynamic graph of related MemCells.- Example: If you are a developer, you might have a MemScene for “Python Project X,” another for “Job Interview Prep,” and another for “Personal Hobbies.”
- Function: When you ask a question about Project X, the system loads the “Project X” MemScene, bringing all relevant history, facts, and files into the Agent’s working memory.
Semantic Consolidation
MemScenes are created and maintained through a background process called Semantic Consolidation. This process runs asynchronously (often when the user is not active), similar to how humans consolidate memories during sleep.1
Clustering
The system analyzes new MemCells using semantic vector embedding. It identifies which existing MemScene the new memory belongs to, or creates a new cluster if the topic is novel.
2
Synthesis
Within a MemScene, the system looks for redundancies or contradictions.
- Redundancy: “I like coffee” and “I drink coffee every day” might be merged.
- Contradiction: “My dog is 3 years old” (old memory) vs “My dog turned 4 today” (new memory). The system updates the facts accordingly.
3
User Persona Update
Key traits extracted from the MemCells are promoted to the global User Profile. This is a high-level summary of who the user is (e.g., “Senior Python Developer,” “Living in New York,” “Prefers concise answers”).
The “Context Window” Problem
LLMs have limited context windows. Even with 1 million tokens, dumping raw logs is inefficient and noisy. MemScenes solve this by providing a pre-digested, structured view of a topic. Instead of reading 500 pages of chat logs to understand “Project X,” the Agent reads the “Project X” MemScene, which contains the synthesized narrative, key facts, and current status.Dynamic Evolution: MemScenes are never finished. They grow and change as you interact with the system, ensuring the AI’s understanding of a topic evolves just as yours does.