Core Concepts
AI Memory Infrastructure
Understand how Omni captures and updates fact-based memories.
AI Memory forms the core foundation of personalized agentic interactions. Traditional systems often retrieve entire chat histories, creating high token costs and distracting the LLM with irrelevant details.
Omni extracts fact-based memories from continuous text logs, updates them dynamically, and organizes them under specific session and developer scopes.
The Lifecycle of a Fact
When user input is sent to the ingest pipeline:
- Ingest & Cleanse: Raw text is parsed to extract key insights while filtering noise, boilerplate text, and temporary comments.
- Fact Consolidation: The pipeline cross-references new facts with existing records inside the user's R2 enclave.
- Upsert or Prune:
- If a fact is new, it is appended to the archive.
- If a new fact contradicts or updates an existing record, the old record is pruned/overwritten (e.g., "User is traveling to Tokyo next week" updates to "User is currently in Tokyo").
Fact Schema
Each memory record contains:
id: A unique UUID.content: The plaintext statement (e.g., "Likes dark mode interface").category: Semantic tag grouping (e.g.,decision,preference,fact).timestamp: Creation and modification times.