Skip to main content
At the heart of EverMemOS lies the MemCell—the fundamental, atomic unit of memory. Unlike a raw text chunk used in vector databases, a MemCell is a structured, semantic object that encapsulates a complete “moment” of interaction.

​
Anatomy of a MemCell

A MemCell is not just text; it is a tuple containing four critical components: M=(E,A,F,T)\mathcal{M} = (E, \mathcal{A}, \mathcal{F}, T)

Episode (E)

A concise, narrative summary of “what happened.” It captures the flow of events and user intent, preserving the causal logic of the interaction.

Atomic Facts (A)

A set of discrete, verifiable statements derived from the episode (e.g., “User likes spicy food,” “Project deadline is Friday”). These facilitate precise factual queries.

Foresight (F)

Forward-looking inferences and predictions. If a user says “I’m flying to Paris tomorrow,” the Foresight component records the future implication (User will be in Paris) with a validity interval.

Metadata (T)

Contextual grounding, including timestamps, location, source confidence, and emotional valence. This helps the system understand when and where a memory applies.

​
From Raw Logs to MemCells

The process of Episodic Trace Formation converts raw chat logs into MemCells.
  1. Segmentation: The system detects boundaries in the conversation (e.g., a topic change or a significant time gap).
  2. Extraction: An LLM parses the segment to extract the narrative (EE) and facts (A\mathcal{A}).
  3. Inference: The system analyzes the text for future implications (F\mathcal{F}).
  4. Packaging: All components are wrapped into a MemCell and assigned a unique ID.
{
  "memcell_id": "mc_123456789",
  "episode": "The user discussed plans for the Q3 marketing campaign, emphasizing a focus on social media channels.",
  "atomic_facts": [
    "User role is Marketing Manager",
    "Q3 campaign focus is Social Media",
    "Budget approval needed by Friday"
  ],
  "foresight": {
    "prediction": "User will submit budget proposal",
    "valid_after": "2023-10-10T09:00:00Z",
    "valid_until": "2023-10-13T17:00:00Z"
  },
  "metadata": {
    "created_at": "2023-10-09T14:30:00Z",
    "source": "slack_integration"
  }
}

​
The “Engram” of AI

Think of a MemCell as an engram in the biological brain—a physical trace of a memory. It bridges the gap between low-level raw data (pixels/tokens) and high-level abstract reasoning. By storing memories as MemCells, EverMemOS ensures that information is:
  • Retrievable: By semantic vector or specific fact.
  • Understandable: Self-contained and context-aware.
  • Actionable: Includes implications for future behavior.