Work architected at Fifth9

Cortex: memory that keeps context and purpose together

Cortex separates capturing information, preserving its context, retrieving relevant records and caring for the resulting memory. Memory retrieval provides evidence to a task; it is a different mechanism from changing the parameters of a neural model.

Memory architecture and implementation work The explanatory layers group responsibilities. They do not claim that every proposed memory tier or learning loop is operational.

Download this guide ↗Return to the story ↗

Cortex layers, recall and learning boundaries ↗ · Original six-tier memory diagram ↗

The detailed view includes additional layers and ownership boundaries; the original six-tier graphic retains its historical scope.

The design question

What should an AI collaborator remember when the work continues?

Help a continuing collaboration recover what matters, where it came from and why it was retained.

Useful memory is more than storage. A retained statement needs enough context to answer who it concerns, where it came from and whether it still applies. Retrieval should help a collaborator inspect relevant evidence, not promote any remembered sentence into unquestionable fact.

Six memory tiers, different responsibilities

Read down from working context to episodes, associations, semantic knowledge, longer-term records and reusable procedures. Each layer has a different purpose and lifetime.

Six-tier Cortex source snapshot: Pulse working context, episodic records, associative graphs, semantic archive, deep archive and procedural Crystal. Storage and lifetimes differ; outcome gating for procedural skills is not proven. — Source architecture snapshot captured on 2 November 2026. It retains that design’s tier names and storage choices; it is not a current deployment inventory. In particular, the diagram explicitly leaves outcome-gated procedural learning unproven. Open the full-size diagram ↗

A view of the responsibilities

Retain, retrieve and reconsider useful context

01Capture with contextSource and scope
02Place it appropriatelyRecord lifetime
03Retrieve for a taskScoped retrieval
04Use and examineEvidence interpretation
05Correct and maintainMemory care

Across the flow

Keep context, responsibility and the boundary between a proposal and an accepted result visible.

The explanatory layers group responsibilities. They do not claim that every proposed memory tier or learning loop is operational. Select a step to read its walkthrough.

What each part is responsible for

The building blocks

Capture and scope

Record relevant information with its intended scope.

Receives: Conversation, work events and operator contributions.
Produces: Scoped records with their origin and purpose.

Working and persistent memory

Support immediate context and records that outlast a session.

Receives: Records with different uses and lifetimes.
Produces: Working context and retained material for later retrieval.

Provenance and lineage

Keep the origin and development of information inspectable.

Receives: Source references, revisions and relationships.
Produces: A record of where information came from and how it changed.

Retrieval

Find material relevant to the current question.

Receives: A question, context and access scope.
Produces: Candidate evidence with enough context to interpret it.

Memory care

Manage correction, retention and removal over time.

Receives: Corrections, retention choices and memory-quality observations.
Produces: Updated records and a deliberate memory lifecycle.

Following one piece of work

How the flow fits together

01Capture with context

Retain a contribution with its purpose, origin and enough context to understand what it refers to. A useful memory should remain distinguishable from a model-generated guess or an observation about a different situation.

02Place it appropriately

Distinguish the context needed for an immediate task from records intended to remain useful later. The explanatory layers describe different lifetimes and responsibilities, not a promise that every proposed memory tier is operational.

03Retrieve for a task

Ask what information this task needs and retrieve within the relevant scope. A semantically similar record can still be outdated, incomplete or inappropriate for the current question.

04Use and examine

Interpret the evidence and check whether it actually helps answer the question. The person or reviewing process still needs to judge whether the context supports the recommendation; retrieval quality and answer quality are separate questions.

05Correct and maintain

Reconsider mistaken or obsolete context and decide what should remain available. Retention, correction and forgetting need an account of derived records as well as original material, and their effects should be examined explicitly.

Illustrative example · not a benchmark result

Return to an earlier design decision

A team returns to a design decision after several weeks.

The collaborator retrieves the earlier requirement, decision and reason for rejecting an alternative. The person can correct an outdated assumption before the next recommendation is made.
If an older decision conflicts with a newer constraint, the person needs to see that conflict and correct the context before relying on the next recommendation.

The useful result: Relevant earlier context that the person can inspect and correct.

Illustrates useful continuity, not a guarantee of recall accuracy or an already-proven outcome-learning loop.

Why these responsibilities are separate

Memory care and learning have different jobs. Correcting a record, changing retention or improving retrieval does not necessarily change a model’s learned capabilities. The architecture makes room for those processes while leaving each improvement claim to an appropriate evaluation.

Reading the architecture accurately

Scope and status

What the work covers

Persistent memory records, scoped retrieval and provenance-oriented archive work.
Benchmark and memory-care work that examines retrieval and retention behaviour.

What still needs to be established

Does retrieved context improve the person’s task?
Which records deserve continued retention?
Can an outcome improve future recall without amplifying mistaken information?

Where the design must hold up

Boundaries and evaluation

Boundary	Why it matters	How to examine it
Retrieval versus answers	Retrieving relevant evidence does not guarantee a correct final answer.	Assess retrieval separately from the final answer and inspect the supporting records.
Scope and retention	Retention must preserve scope and permit correction or forgetting.	Test relevant scope boundaries, corrections and the effects of retention choices.
Measured improvement	Recorded outcomes do not prove that memory quality improves because of them.	Compare an explicit memory change against an appropriate baseline instead of inferring improvement from new records.