Forgetting‑First AI

Design Principles and Reference Architecture for Consent‑Aware Organizational AI

Purpose

This document defines a forgetting‑first design philosophy for deploying AI in organizations where consent, boundaries, and agency must remain intact over time.

It complements the Consent‑Aware AI in Organizations taxonomy by translating theory into:

Concrete design principles
A modular reference architecture
Explicit forgetting levers that are technical, not rhetorical

The core premise is simple:

Forgetting must be cheaper than remembering, and safer than compliance.

Part I — Design Principles

These principles are non‑optional if real forgetting is desired. Violating any one of them will reintroduce silent memory accumulation.

Principle 1: Indirection Before Intelligence

Never expose first‑order meaning to AI when second‑order structure is sufficient.

Replace identifiers with opaque tokens
Abstract sensitive values into buckets, ranks, or classes
Preserve relational structure without referents

Forgetting lever: destroy the indirection index

Principle 2: Context Is a Boundary, Not a Prompt

Context must be enforced structurally, not requested linguistically.

Context is scoped by role, purpose, and time
Cross‑context access requires explicit re‑witnessing

Forgetting lever: expire or delete context shards

Consent must be independently scoped for:

Contribution
Inference
Learning
Decision authority

Consent in one loop does not propagate to others.

Forgetting lever: revoke consent → automatic non‑propagation

Principle 4: Memory Must Be Costed

Persistence is never free.

Default state is non‑persistence
Storage requires justification, witnessing, and scope

Forgetting lever: unattended memory decays automatically

Principle 5: Time Is a First‑Class Constraint

All AI‑touched data must carry an expiration horizon.

Different loops decay at different rates
Renewal requires renewed consent

Forgetting lever: time‑based invalidation

Principle 6: Witness Before Canon

No AI output becomes organizational memory, policy, or training input without a human witness taking epistemic ownership.

Witnessing is not approval; it is accountability.

Forgetting lever: unwitnessed outputs evaporate

Principle 7: Learning Is Air‑Gapped

Inference systems and learning systems must be separated.

Most interactions should never affect model behavior
Learning occurs slowly, deliberately, and audibly

Forgetting lever: inference models are disposable

Principle 8: Similarity Is Scoped

Embedding spaces encode memory implicitly.

Separate embeddings by role, purpose, and consent domain
Do not mix vectors across boundaries

Forgetting lever: delete embedding spaces, not just records

Principle 9: Non‑Optimization Is a Feature

Some domains must remain intentionally under‑optimized.

HR
Governance
Conflict resolution

What is never learned never needs to be erased.

Part II — Reference Architecture

This architecture is conceptual, not vendor‑specific. It describes control surfaces, not implementation details.

1. Boundary Layer (Pre‑AI)

Purpose: enforce indirection and scope before AI contact

Components:

Identity tokenizer
Sensitive value abstraction
Role + purpose scoping

Outputs:

Opaque tokens
Structured, non‑identifying representations

2. Context Shard Manager

Purpose: prevent context bleed

Responsibilities:

Create per‑task, per‑role context shards
Enforce shard isolation
Track shard expiration

Failure mode prevented: cross‑role inference reuse

3. Inference Engine (Stateless)

Purpose: generate outputs without memory

Characteristics:

No long‑term state
No self‑learning
Disposable instances

Critical constraint: outputs are non‑canonical by default

4. Inference Classification & Tagging

Purpose: limit propagation

Each output is tagged with:

Inference class (descriptive, predictive, evaluative, speculative)
Allowed downstream domains
Expiration horizon

5. Witness Gate

Purpose: control transition from artifact → attractor

Function:

Human explicitly witnesses output
Confirms contextual validity
Accepts accountability

Absent witness: output cannot persist

6. Memory Layer (Lossy by Design)

Purpose: store only what must persist

Constraints:

No raw transcripts by default
Summary‑only storage
Template‑enforced compression

Storage objects include:

Consent scope
Expiry
Witness ID

7. Learning Pipeline (Air‑Gapped)

Purpose: deliberate system improvement

Inputs:

Curated, witnessed, consented summaries

Controls:

Audit trails
Slow update cadence
Rollback capability

8. Expiry & Forgetting Engine

Purpose: make forgetting automatic