Core Concepts

Understand Hebbrix's storage boundaries, operational memory tiers, temporal truth, and provenance without relying on brain metaphors.

01. Memories

A "memory" is just a piece of information. It could be:

A chat message ("User prefers dark mode")
A document you uploaded (your company's FAQ)
A preference ("Speaks Spanish")
Literally anything you want the AI to remember

02. Collections

Collections are folders for organizing memories. Like:

customer-support - All support conversations
product-docs - Your documentation
user-preferences - Settings and choices

When searching, you can filter by collection to find specific types of memories faster.

03. Operational Memory Tiers

Hebbrix can organize memory into three operational lifecycle tiers:

Short-Term Memory (STM)

Recent, task-local context that is useful for active work and short horizons.

Example: "User asked about pricing 2 hours ago"

Medium-Term Memory (MTM)

Reusable context whose access, recency, and importance justify broader retrieval.

Example: "User's name is John, prefers email communication"

Long-Term Memory (LTM)

Durable knowledge intended for long-lived retrieval, still subject to correction, supersession, retention, and explicit deletion.

Example: "Our refund policy is 30 days", "John is a premium customer"

Tiers influence lifecycle and ranking; they are not promises that a fact is true forever or that unused data is silently deleted. Temporal validity, corrections, supersession, retention policy, and explicit deletion remain authoritative.

04. Knowledge Graph

Behind the scenes, Hebbrix builds a graph of how things connect. For example:

Knowledge graph

John → works_at → Acme Corp
John → prefers → email
Acme Corp → industry → SaaS
email → faster_than → phone_calls

This supports relationship and point-in-time queries. Smart ingestion can extract graph evidence automatically; manual entity, relationship, query, timeline, and contradiction APIs provide control and auditability.