***

title: Platform overview
description: Architecture and capabilities of the XTrace memory platform.
-------------------------------------------------------------------------

# Platform overview

XTrace is a managed memory platform with three layers:

## Extraction

When you send conversation turns to the API (or sync them via the Chrome extension), XTrace runs a **3-way parallel extraction pipeline**:

* **Beliefs (facts)** — atomic assertions about the user, their preferences, decisions, and context
* **Episodes** — structured summaries of what happened in a conversation session
* **Artifacts** — work products (code, documents, strategies) with version history

Extraction uses specialized prompts for each type, with post-hoc linking between facts and the episodes they came from.

## Revision

Extracted knowledge isn't static. The **belief revision engine** continuously maintains consistency:

* **Supersession** — new beliefs replace outdated ones, preserving lineage
* **Contraction** — beliefs can be retracted without replacement ("we're NOT doing X")
* **Entrenchment** — user-stated beliefs outrank system inferences; core beliefs are protected
* **Consolidation** — duplicate and near-duplicate facts are merged via text dedup + LLM resolution

## Retrieval

When you search or an agent requests context, XTrace assembles the most relevant knowledge:

* **Semantic search** across beliefs, episodes, and artifacts
* **Classifier-gated routing** to the right knowledge type
* **Iterative reranking** for precision
* **Character-budgeted context assembly** — fits the most relevant information into your token budget

## Security

All data is stored in XTrace's encrypted vector database. Each user's knowledge is isolated in scoped, permission-controlled stores. API keys are generated per-user with org-level scoping.