One place for everything about one thing.

A topic is how MemState bundles everything learned from agent observations about one subject — identity, summary, fields, importance, and links. Context queries return that bundle as a coherent unit instead of scattered rows.

What stays in sync

MemState keeps three concerns together so retrieval stays honest as the agent learns.

Content

The facts themselves: titles, summaries, typed fields, and the text used to understand meaning.

Structure

How topics relate: parent-like links, peer associations, and references embedded in fields.

Policies

The rules that cap history length, decide what is current, and move stale material out of the default view.

What a topic graph is (picture)

Think of memory as cards (each card is one topic) and strings between cards when the agent learned a relationship. That whole picture is the topic graph: rich nodes, typed links, optional pointers from a field on one card to another.

The topic graph is that web of cards and links. MemState keeps each card self-contained and, when the agent asks for context, loads the cards that matter and optionally expands along links.

How this improves memory quality

Because each topic is self-contained, one context response can carry the whole picture. Because structure is typed, expansion along relationships is predictable. Because policies are centralized inside MemState, the agent loop does not re-implement forgetting and supersession per integration.

How to use it in product terms

Start a new topic when the agent learns a new subject that will be referenced again.
Add fields as the story grows instead of overwriting earlier facts.
Link out when another subject should stay its own topic with its own history.

Continue

Field history — why past values stay available.
Links between topics — how context expands safely.
Adding memory — the write path in plain language.