Skip to content

introduction.md

Latest commit

 

History

History
62 lines (35 loc) · 3.43 KB

introduction.md

File metadata and controls

62 lines (35 loc) · 3.43 KB
Raw

Atomic Memory, explained

Agent index: llms.txt

An AI model can use the words in its current context window. Memory lets your application carry useful context into the next conversation — deliberately, within the right boundary, and with a record you can inspect later.

Atomic Memory is managed memory for that job. Your application writes a fact in the right scope, retrieves it for the next task, corrects it when it changes, and can inspect every step. You call an API; there is no engine to operate.

Start with a complete loop

The Quickstart writes a memory, retrieves it in a later request, and lets you confirm both operations in the console.

What durable memory changes

Imagine a travel assistant. In one chat, a person says: “I prefer aisle seats when I fly.” On a later day, the person asks which seat to book. The assistant should not have to guess, or rely on an old chat still fitting in the model’s context window.

Memory gives the application a deliberate path between those moments:

How one preference becomes durable memory — and changes safely over time.

  1. 1. Conversation“I prefer aisle seats when I fly.”
  2. 2. ScopeKeep it with the right person: user_001.
  3. 3. Mutation decisionAdd, update, delete, or leave the record unchanged.
  4. 4. Durable memorySeat preference: aisle
  5. 5. Retrieve“Which seat should I book?”
  6. 6. Assistant response“I’ll look for an aisle seat.”

Source: chatVersion: v2Changed: just now

Earlier recordv1 · Aisle seat

→

Corrected recordv2 · Window seat

A correction preserves the history instead of silently overwriting it.

The diagram is a mental model, not hidden magic. Your application chooses the scope, Atomic Memory records the memory and its lineage, and retrieval returns useful context for the next model call. If the preference changes, the new record can supersede the old one without erasing the reason it changed.

Three ideas to carry with you

Scope keeps memory in the right place

A scope is the boundary a memory belongs to. Most applications begin with a user scope: one person’s preferences should not appear in another person’s conversation. Teams can add workspace or agent boundaries as their product needs them. Learn more in Scopes and identity.

Retrieval brings back only what helps

Retrieval searches the relevant scope for context that can help answer the current request. It is not a transcript dump: the point is to give the model a small, relevant starting point for the next response.

Provenance keeps memory accountable

Provenance is the trail back to where a memory came from and how it changed. It is what lets you inspect a surprising answer, correct a bad fact, and understand which version is active.

What you get

  • Console visibility — every ingest, search, and mutation writes an inspectable trace with its call type, timing, and reconciliation decision. Memory Explorer shows the actual stored content.
  • Portable integrations — the same SDK, MCP tools, and HTTP API back your agents and frameworks. See Integrations.
  • Nothing to operate — no database to run, no engine to upgrade.

Next step

Run the Quickstart. You will create one memory, retrieve it, and see where to inspect it before moving on to the full memory lifecycle.