Skip to content

ENTERPRISE MEMORY / A QUESTION OF REUSE

When may an agent
reuse a memory?

Move from a meeting suggestion to an approved decision, then see what changes when evidence conflicts or access is revoked.

REMEMBERING IS
ONLY THE BEGINNING.

Correct.
Current.
Permitted.

A memory needs all three before it can earn reuse.

Scripted demonstration using synthetic data. No live customer records, memory service or security enforcement is connected.

MEMORY REUSE EXPLORERSYNTHETIC SCENARIO / 18 NOVEMBER 2026
01 / WHO IS ASKING?
02 / WHAT CHANGED?Choose the evidence state

Approval is a teaching state you select. It is never generated by a model. Dates belong to this synthetic scenario.

THE SHARED KNOWLEDGE WORKSPACE

An organisation that remembers.
And knows what its knowledge means.

Follow people and agents as they read, contribute and review. Every record has an owner, a type and a history. Every connection has a meaning.

43RECORDS06HUMANS06AGENTS64RELATIONSHIPS
SIMULATION PAUSEDScene 01 / 18 · replay resets to baseline
SIX DEPARTMENTS / SHARED CONTEXTSELECT A NODE TO INSPECT ITS HISTORY
100%
Choose a department above to see every label.
HUMAN / PEOPLE / CONTRIBUTE

Aisha updates joiner readiness

Human contribution: adds a named Technology handoff before access is enabled.

Aisha CONTRIBUTES_TO Joiner readiness
Human ownerSpecialist agentReadingContributingBlocked or rejected
SELECTED RECORD / SHARED

Shared knowledgeKnowledge · v2

Versioned records connected by explicit business relationships. Ownership, source evidence and current permissions travel with the knowledge.

Accountable owner
Department owners
Current state
KNOWLEDGE
Stable record ID
memory
Type determines which relationships are valid. Ownership, approval and permissions are checked separately.

Version history 2 versions

New contributions preserve history. A proposed revision does not grant approval.

v2 / Before this replay

Typed relationships and source context reviewed.

Department owners · Human

Typed relationships 6

These are business facts with defined types, not just visual links.

ONTOLOGY IN ACTION

A shared vocabulary.
Rules that can be checked.

Entities have stable identities and types. Relationships specify which types may connect. Versions preserve how those facts change.

Current activity: types valid

Human → CONTRIBUTES_TO → Process. A contribution creates a version; it does not grant approval.

Decision

Relationship rejected

APPROVES requires Human → Decision or Policy. Agent → Decision is rejected. This tests the vocabulary only; it does not change any record or permission.

Drag nodes to stretch their connections; release to let them settle. Drag the canvas to pan. Use + / −, Ctrl + wheel or pinch to zoom. Keyboard: Tab to a node, Enter to inspect, arrow keys to pan, 0 to reset. Synthetic organisation and scripted activity; no live services or customer records.

03 / MAY THIS MEMORY BE REUSED? No recommendation to proceed

Proposed

Visible evidence

Sample meeting note

A participant suggests SSO-only access for the North customer.

Proposed response

No approved configuration change. Ask the owner to validate the source.

Why this result?

A suggestion has not been accepted as an implementation instruction.

Scripted demonstration using synthetic data. No live customer records, memory service or security enforcement is connected.

A PROPOSED DESIGN

Experience becomes knowledge
through deliberate checks.

Confidence is not authority. A model may propose a claim; its confidence does not grant permission or make the claim true.

  1. 01

    Raw source

    Retain the original record and its permissions.

  2. 02

    Candidate claim

    Extract a possible fact. Model extraction can be wrong.

  3. 03

    Validation

    Check evidence, entity matches and conflicting claims.

  4. 04

    Approved memory

    A trusted owner approves the scope and permitted audience.

  5. 05

    Authorised retrieval

    Check current permissions and valid dates for this task.

  6. 06

    Observed outcome

    Record what happened and any correction required.

  7. 07

    Revalidation

    Revisit the memory when sources, permissions or outcomes change.

Keep the context with the claim.

Derived memory needs source lineage, scope, owner, valid dates, recorded dates, approval status and audience. A note-link graph and a temporal knowledge graph serve different purposes.

Make changes travel.

Revocation, erasure and source updates must invalidate relevant derived records, caches and summaries. This is a design requirement, not something this demo or any product supplies automatically.

READ THE DECISIONS

Every state has
a different consequence.

Meeting suggestion

A proposed claim does not authorise an implementation change.

Approved change

The scoped owner may use the current approved instruction and cite its evidence.

Conflicting evidence

Reconcile apparently authoritative records before making a recommendation.

Superseded decision

Keep the history for explanation. Retrieve the replacement under current permissions.

Access revoked

Withhold the source and derived answer. Previously cached context must also be invalidated.

Untrusted instruction

Quarantine content that tries to approve itself. Source text cannot grant authority.

In every state, an actor outside the source’s scope receives neither its content nor a derived answer.

THE THINKING BEHIND THE DEMONSTRATION

From Agent Memory
to Enterprise Memory.

Read the research, the proposed architecture and the questions worth testing in your own organisation.

Read the article