AI Memory Needs a Forgetting Boundary: Persistent Context Is a Governance Decision
Do not switch on persistent AI memory as a blanket productivity setting. Separate project context, personal preferences, and approved organizational playbooks; give each a purpose, owner, access boundary, retention period, correction route, and tested deletion procedure. Memory can reduce repeated setup, but it does not prove that an old instruction is still authorized, accurate, lawful, or appropriate for a new market. A brand should scale memory only after it can show what the system remembers, where that memory is used, and how it is reliably forgotten.
Do not switch on persistent AI memory as a blanket productivity setting. Separate project context, personal preferences, and approved organizational playbooks; give each a purpose, owner, access boundary, retention period, correction route, and tested deletion procedure. Memory can reduce repeated setup, but it does not prove that an old instruction is still authorized, accurate, lawful, or appropriate for a new market. A brand should scale memory only after it can show what the system remembers, where that memory is used, and how it is reliably forgotten.
What the launch actually establishes—and what remains a claim
Harvey announced on August 18 that its legal agents can open inside a matter or project with documents, parties, tasks, permissions, and work history already present. It also says a personal memory can carry writing preferences across Harvey, Word, and Outlook; users can view, change, or turn that memory off. These are verified descriptions of the announced product design. The further statement that more context lets users trust an agent with more substantive work is Harvey's product claim, not independent proof of accuracy, legal quality, or business value. The event matters because it makes a broader application shift concrete: AI is no longer receiving only a prompt. It is being given an inherited operating history.
- Verified fact: the announcement is dated August 18, 2026 and falls inside this article's 72-hour window.
- Source claim: persistent context and preferences reduce repeated setup and raise the ceiling on delegated work.
- ZEHUA judgment: less setup is useful, but inherited history also inherits stale assumptions and access risk.
One word, three different memory contracts
Brands should not manage all remembered material as one pool. Project context contains the documents, people, approvals, claims, offers, and tasks for a particular campaign or market. Personal preference memory contains how an individual structures a summary, cites evidence, or edits language. Organizational playbooks contain approved terminology, escalation rules, claim boundaries, and reusable formats. These categories can overlap, but they should not silently merge. A preference learned from one employee must not become corporate policy; a client or market instruction must not leak into another workspace; and an approved playbook must not be overwritten by a convenient correction from a single task. The three categories need separate owners, permissions, and expiry logic.
- Project context should remain inside the named matter, market, or campaign boundary.
- Personal preference memory should be inspectable, correctable, portable only by choice, and removable.
- Organizational playbooks need version control and named approval, not passive learning alone.
Build a memory register before the pilot becomes infrastructure
For every memory-enabled workflow, record seven fields before launch: purpose, data category, source, owner, allowed destinations, expiry trigger, and deletion proof. Add a decision for conflicts: when a current project instruction disagrees with a remembered preference or an older playbook, which source wins? Harvey's security page says customers can set retention, delete data, enforce role-based access, and keep workspaces logically separated. Its service terms also warn that some optional features can have different retention, human-review, or processing-location requirements. Those documents should be read together. A procurement checklist that asks only whether data trains a model misses the operational question of what persists at inference time, who can retrieve it, and which feature-specific terms change the answer.
- Name the system of record for permissions, playbooks, and deletion requests.
- Test retrieval and deletion with a harmless canary fact before using sensitive work.
- Recheck feature-specific terms whenever browsing, collaboration, email, or extended AI functions are enabled.
The boundary: when memory may scale, and when it should stop
Scale persistent memory only when users can inspect and correct it, administrators can restrict its destinations, retention matches the purpose, deletion is testable, and a named human still approves consequential output. Pause when the system cannot explain which layer supplied an instruction, when a person changes role, when a campaign or client ends, when consent or contractual scope is unclear, or when data would cross a market or workspace boundary. China's Interim Measures apply to public generative-AI services in the stated scope and require protection of user inputs and usage records, avoidance of unnecessary personal-information collection and unlawful retention, and handling of access, correction, and deletion requests. They are not a universal rule for every internal tool, but they reinforce a practical operating principle: remembering more is not automatically safer or more valuable. This framework is operational guidance, not legal advice.
- What changed: a current product launch makes inherited project and personal context a visible application pattern.
- What did not change: memory does not renew authorization, evidence, rights, accuracy, or human accountability.
- Next step: run a 14-day pilot with one memory class, one owner, one expiry rule, and a witnessed deletion test.
Questions a serious decision should answer.
Short answers first, with the boundary made visible.
Is turning off model training enough to make AI memory safe?
No. No-training commitments address one risk. A brand must still understand inference-time retention, workspace access, subprocessors, feature-specific exceptions, correction, deletion, and whether remembered instructions remain authorized.
Should an AI remember a brand voice across every market?
Usually not as one global memory. Keep a governed core playbook, then attach market-specific language, cultural, claim, and legal layers with separate owners and expiry. A correction from one market should not silently rewrite another.
What is the smallest useful memory pilot?
Use a reversible, low-risk workflow such as internal summary formatting. Seed one harmless preference, confirm where it appears, change it, revoke access, delete it, and verify that it no longer returns before adding sensitive context.
Four languages. One place to exchange what the market is really saying.
Opening the exchange…
Write without creating an account.
A public display name is enough to submit a comment or reply. Sign in only when you want an avatar, saved topics, helpful reactions, or reporting tools.
Sign-in is temporarily unavailable. Guest comments still work.
Your public avatar
Choose one of eight ZEHUA signals, or upload a photo. Uploads are cropped locally and saved as a 160×160 image.
Bring the real constraint. We will map the next route.
No fixed package before the question is understood.