Memory for AI systems, kept across models.

Cortex records what a system knows, where each fact came from, and how far to trust it.

What Cortex includes

Memory, provenance, and context

Three modules, and each can say where a fact came from.

Engram

Building

Memory

Engram keeps a store of what a system has seen and a store of what it has learned, and moves facts from the first to the second.

Trace

Building

Provenance

Trace links every memory, context, and output to its inputs, so you can check where any claim came from. It is the core of Cortex.

Recall

Declared

Context

Recall chooses which known facts to give a model for the task at hand. It is planned, and work has not started.

Trace

Every answer, traced to its inputs

Trace links every output to what it was derived from. Choose an answer, and the drawing lights everything it came from.

Read from left to right: what came in, the facts Engram kept from it, what was retrieved and looked up, the context the model was given, and the answer it gave. The chosen answer is lit back to where it came from. An example of four answers from a support assistant, built by Trace and Engram and drawn with Helix Graph. Trace is still being built.

“Shipped on the 8th.”

“Pro is $15 a seat.”

“Pro is $20 a seat.”

“Within 30 days.”

What “Pro is $20 a seat.” came from

ContextsOctober pricing context
Tool callssearch_docs, lookup_account
MemoriesOctober price retrieval
Facts, in EngramPro is $20 a seat, On the Pro plan
SourcesWhat will my plan cost?, October pricing, Support chat

Delete “October price retrieval”

Deleting a memory removes everything derived from it, as Trace does. The September answer cites the superseded price, which is why Engram keeps it.

Why a memory layer

The record stays when you replace the model

Cortex records what a system knows, where it learned it, and how far to trust it. That mattered less when a person wrote every line. It matters now that most lines are generated.

What it knows

Engram keeps what a system has seen and what it has learned, as two stores, and moves facts from the first to the second.

Where it learned it

Trace links every memory, context, and output to its inputs, so you can follow any claim back to the source it came from.

How far to trust it

The record is kept apart from the model. Replace the model, and what the system knows and where it came from stays.

Work with us

We take on work that also improves Cortex, so each engagement adds to the foundation.

Start a conversation

BuildingEngram and Trace are underway

Memory for AI systems, kept across models.

Cortex records what a system knows, where each fact came from, and how far to trust it.

See an example trace Read the thesis

What Cortex includes

Memory, provenance, and context

Three modules, and each can say where a fact came from.

Engram

Building

Memory

Engram keeps a store of what a system has seen and a store of what it has learned, and moves facts from the first to the second.

Trace

Building

Provenance

Trace links every memory, context, and output to its inputs, so you can check where any claim came from. It is the core of Cortex.

Recall

Declared

Context

Recall chooses which known facts to give a model for the task at hand. It is planned, and work has not started.

Trace

Every answer, traced to its inputs

Trace links every output to what it was derived from. Choose an answer, and the drawing lights everything it came from.

“Shipped on the 8th.”

“Pro is $15 a seat.”

“Pro is $20 a seat.”

“Within 30 days.”

What “Pro is $20 a seat.” came from

ContextsOctober pricing context
Tool callssearch_docs, lookup_account
MemoriesOctober price retrieval
Facts, in EngramPro is $20 a seat, On the Pro plan
SourcesWhat will my plan cost?, October pricing, Support chat

Delete “October price retrieval”

Deleting a memory removes everything derived from it, as Trace does. The September answer cites the superseded price, which is why Engram keeps it.

Why a memory layer

The record stays when you replace the model

Cortex records what a system knows, where it learned it, and how far to trust it. That mattered less when a person wrote every line. It matters now that most lines are generated.

What it knows

Engram keeps what a system has seen and what it has learned, as two stores, and moves facts from the first to the second.

Where it learned it

Trace links every memory, context, and output to its inputs, so you can follow any claim back to the source it came from.

How far to trust it

The record is kept apart from the model. Replace the model, and what the system knows and where it came from stays.

Work with us

We take on work that also improves Cortex, so each engagement adds to the foundation.

Start a conversation

Docs

Tools

Memory for AI systems, kept across models.

Memory, provenance, and context

Every answer, traced to its inputs

The record stays when you replace the model

Work with us

Docs

Tools

Memory for AI systems, kept across models.

Memory, provenance, and context

Every answer, traced to its inputs

The record stays when you replace the model

Work with us