# The Memory Model

Cortex is a memory system with structured artifacts and durable state, not a chatbot
wrapper or a prompt library. Five artifact types form its backbone. They live in
`backend/mcp/src/core/models.ts`.

## The artifacts

| Artifact | What it is |
| --- | --- |
| **`Source`** | A raw input: note, document, image, audio, video, URL, or structured payload |
| **`Memory`** | A durable extracted memory with text, tags, confidence, provenance, and timestamps |
| **`Extraction`** | The result of processing a source into memories |
| **`MemoryDiff`** | A structured consolidation result that can merge, verify, prune, or pattern-match |
| **`NamespaceManifest`** | The per-namespace pointer record for versions and artifacts |

## How they relate

```text
Source ──(extract)──▶ Extraction ──▶ [ Memory, Memory, … ]
                                            │
                                  recorded in
                                            ▼
                                   NamespaceManifest  ◀── MemoryDiff (consolidation)
```

::::steps

##### A Source enters

Raw, untouched input. The original is preserved so memory always has provenance to trace
back to.

##### Extraction produces Memories

Processing a source yields an `Extraction` containing the durable `Memory` records it
distilled, each carrying tags, a confidence score, and a link back to its source.

##### The Manifest tracks state

The `NamespaceManifest` is the namespace's source of truth: it points at the current
versions and artifacts, so the head is always well-defined and history is traceable.

##### A MemoryDiff evolves it

Consolidation emits a `MemoryDiff` describing how memory should change (merges,
verifications, prunes, pattern matches), which is then applied to advance the manifest.

::::

## Memory is first-class

Because memories are structured records rather than chat lines, Cortex can do things a
transcript can't:

:::tip\[Provenance & confidence]
Every memory knows where it came from and how sure Cortex is about it. Recall is
**verified-first**, so trusted facts surface ahead of unconfirmed ones.
:::

* **Tags** make memory navigable and group-able.
* **Timestamps** give every memory a position in the [timeline](/concepts/how-cortex-works).
* **Tombstones** let you [forget](/concepts/sharing) a memory without destroying its raw record; recall simply stops surfacing it.

## Derived views

From these artifacts, Cortex produces read-only **derived views** without changing state:

| View | Answer it gives |
| --- | --- |
| **Tags** | What themes dominate this namespace? |
| **Digest** | What does a period of memory summarize to? |
| **Timeline** | How has the namespace changed across versions? |
| **Connections** | Which memories relate to one another? |

These power the app's memory surfaces and are exposed as MCP tools. See the
[Tool Reference](/mcp/tools).

## Keep reading

* Where these artifacts physically live: [Infrastructure](/infrastructure/overview).
* How the diff keeps memory clean: [The Elastic Brain](/concepts/elastic-brain).