Skip to main content
Traditional search is passive: you type a query, and the system finds documents that match your words. Reconstructive Recollection is active: the Agent analyzes your current intent and rebuilds the necessary context from its memory store to best support the task at hand.

​
Beyond Vector Similarity

Standard RAG systems rely heavily on vector similarity search (k-NN). While fast, this often fails in complex scenarios:
  • The “Lost in the Middle” Phenomenon: Retrieving too many loose chunks confuses the LLM.
  • Lack of Context: A retrieved chunk might be factually correct but irrelevant to the current nuance.
EverMemOS uses a dual-process approach:
  1. Scene Activation: First, the system identifies which MemScene (topic) is currently active.
  2. Agentic Walk: The system then traverses the connections within that MemScene to find specific MemCells that answer the query.

​
The Reconstruction Process

When a user prompt arrives (e.g., “Draft an email update for the project we discussed last week”):
1

Intent Analysis

The system determines the user wants to write an email (Task) about “the project” (Topic) referencing “last week’s discussion” (Time/Constraint).
2

Scene Loading

The relevant “Project” MemScene is loaded.
3

Context Synthesis

The system doesn’t just return the raw transcript of last week’s meeting. It reconstructs the memory:
  • It pulls the Decisions Made (Fact).
  • It pulls the Action Items assigned to the user (Fact).
  • It ignores the Small Talk about the weather (Irrelevant Episode).

​
Adaptive Retrieval

Reconstructive Recollection is adaptive.
  • For factual queries (“What is my API key?”), it performs a precise lookup of Atomic Facts.
  • For creative tasks (“Brainstorm ideas based on my previous notes”), it retrieves broader episodic narratives to inspire the model.
  • For reasoning tasks (“Why did we decide to switch databases?”), it traces the causal chain across multiple MemCells to explain the history of a decision.
This mechanism mimics human recall. When you tell a story, you don’t replay a video tape of your life; you reconstruct the narrative based on the key highlights relevant to your current listener.