Skip to main content
EverMemOS operates through a cognitive loop involving two main tracks: Memory Construction and Memory Perception.

​
System Framework

The architecture is designed to turn raw conversation streams into structured, retrievable knowledge.
EverMemOS Architecture

​
Memory Construction Layer

This layer is responsible for ingesting data and organizing it into meaningful units.
1

Ingestion

Raw messages enter the system through the API or data pipeline.
2

Boundary Detection

The system identifies shifts in topics or contexts to segment the conversation.
3

Extraction

Specialized prompts and models extract MemCells (Atomic Memory Units).
4

Consolidation

Integrate by theme and participants to form episodes and profiles.
5

Indexing

Data is stored with both keyword (lexical) and semantic (vector) indices for robust retrieval.

​
Memory Perception Layer

This layer handles how the agent “remembers” and uses information.
  • Intelligent Retrieval Tools:
    • Hybrid Retrieval (RRF): Combines keyword (BM25) and vector search using Reciprocal Rank Fusion.
    • Reranking: A specialized model reorders candidate memories to ensure relevance.
  • Flexible Strategies:
    • Lightweight Fast Mode: Optimized keyword search for low latency.
    • Agentic Multi-Round: Generates clarifying questions for complex queries.
  • Reasoning Fusion:
    • Recalled memories are fused with the current conversation context to prevent hallucinations.