Skip to main content
EverMemOS employs multiple retrieval strategies to balance speed and accuracy.

​
Strategies

Lightweight Fast Mode

Mechanism: Pure Keyword Search (BM25). Use Case: Real-time typing suggestions, low-latency autocomplete.

Agentic Multi-Round

Mechanism: LLM-driven query generation + Multi-step search. Use Case: Complex reasoning, answering “Why” questions.

​
The Retrieval Pipeline

  1. Query Analysis: The system analyzes the user’s input to determine intent.
  2. Query Expansion: Generates multiple search queries (keywords, vectors) to broaden coverage.
  3. Hybrid Search: Executes BM25 and Vector search in parallel.
  4. RRF Fusion: Merges results using Reciprocal Rank Fusion.
  5. Reranking: Scores the fused list to select the top candidates.
  6. Reasoning: The LLM uses the retrieved memories to formulate the final response.