Your Oppy Has a Memory. Give It an Expiration Date Before It Gives a Customer Yesterday's Answer.
By Anna with Oppy
The ten-second thesis: An AI employee should never give yesterday's answer with today's confidence. Give every operational fact an owner, a source, an approved use, and an expiration date.
A property manager asks an AI employee whether a resident may pay by an old method. A loan officer asks which document is required. A title team asks whether a recorded exception was cleared. A dental office asks whether a membership benefit still applies.
The answer may be beautifully written, promptly delivered, and wrong by three policy versions.
This is the next operating problem for agentic AI. Businesses are teaching AI employees from call transcripts, CRM notes, uploaded files, meeting summaries, websites, and connected tools. Most teams can explain how knowledge enters the system. Far fewer can explain when it becomes too old to use.
That gap now matters. Buildium reports that AI adoption among property management companies rose from 20% in 2024 to 58% in 2025, yet only 8% had fully automated any process.1 Adoption is sprinting. Operational maturity is still looking for its shoes.
A September 7 X post from property manager Peter Lohmann supplied the live operator version, quoting Join Crane's new report:
“Property managers are more interested in AI agents at this point than building and maintaining their own software and tools, even if it brings cost savings.”10
That quotation is a timely industry signal, not a prevalence study. It still clarifies the buying question. Operators want the job done. The useful product is not another tool to maintain. It is a managed employee whose sources, limits, and exceptions remain visible.
Build a Knowledge Expiration System: four narrow Oppies and one shared register that make stale knowledge easier to stop than to repeat.
Why memory needs a clock
On September 7, the UK National Cyber Security Centre warned that unapproved AI can reduce an organization's visibility and control over company or customer data. Its summary is admirably short:
“You cannot manage what you do not know.”2
The same logic applies inside an approved AI system. You cannot manage a fact if you do not know where it came from, who owns it, what it may support, or when somebody must review it.
The industry is moving toward more structured access. RISMedia reported that Bright MLS planned a mid-September Model Context Protocol server so subscribers could call approved MLS data when needed rather than download raw files or paste records into external tools.3 This is a useful direction. Query-time access can improve control. It does not answer the full lifecycle question.
A source can be permissioned and still become stale. A transcript can be accurate and still lose authority. A CRM note can be useful for service and still be insufficient permission for a marketing text.
Cotality's first-party international survey of recent and prospective homebuyers in the U.S., Canada, the U.K., and Australia, fielded January 29 to February 9, 2026, found that 64% of respondents were concerned AI might recycle unverified information rather than validated first-party data. It also found that 44% would pay an additional fee for a human expert to verify AI-generated housing decisions.4 These surveyed buyers are expressing a need for safeguards alongside speed.
The missing object: a Knowledge Lifecycle Register
A Knowledge Lifecycle Register is not another document library. It is a small control record attached to every knowledge item an AI employee may use for consequential work.
| Register field | Required answer |
|---|---|
artifact_id |
Which exact document, transcript, note, page, or source record is this? |
authority_source |
Is it an official policy, a system-of-record field, a customer statement, or informal context? |
owner |
Which person or role is accountable for its continued use? |
approved_use |
May it support internal research, a draft, customer service, or an outward-facing message? |
sensitivity |
What is the minimum access needed to handle it? |
effective_at |
When did this version become usable? |
review_at |
When must a human confirm it is still current? |
expires_at |
When must the AI stop using it unless renewed? |
correction_route |
Where does a challenge or new source go? |
retention_rule |
Preserve, archive, redact, request deletion, or hold for authorized review? |
status |
current, review_due, expired, conflict, or needs_attention |
evidence_pointer |
What proves the last review, correction, approval, or disposition? |
The point is not to assign a universal shelf life. A published office holiday schedule may expire in 30 days. A customer-stated preference may remain useful until withdrawn. A title document may need preservation under a business rule that no AI employee should invent. The register forces somebody to make the rule visible.
Build four Oppies with incompatible powers
OpenAI reported on September 6 that, in an internal analysis of coding-agent use by its research organization, more than half of successful tasks estimated at four to eight hours involved at least one human intervention. The company called its measurements preliminary; they concern AI research, not a general benchmark for business agents.5 The operating lesson still travels well: longer agentic work needs deliberate human checkpoints.
Do not build one omniscient memory bot. Build four small employees that cannot quietly approve their own work.
| Oppy | Narrow job | May do | Must not do |
|---|---|---|---|
| Intake Registrar | Create the lifecycle record before knowledge enters an operational workflow | Link the artifact, record the owner and approved use, calculate a review date from an approved rule | Decide that a source is legally authoritative, safe to retain, or suitable for a licensed decision |
| Freshness Checker | Compare registered items with approved source versions before scheduled reviews and consequential use | Flag expiry, unavailable sources, superseding versions, or conflicts | Pick a winner when authorities conflict or silently rewrite source material |
| Correction Clerk | Turn a challenge into a traceable case | Preserve the original pointer, link the correction source, identify affected work, assign a reviewer | Treat an allegation as fact, erase history, or contact a customer without an approved workflow |
| Retention Gate | Prepare a disposition packet for an authorized person | Identify dependencies, holds, minimum audit evidence, and the system of record that owns the action | Delete, redact, archive, or override a hold without human approval |
BCG calls a pre-governed, production-ready starting point a golden path, with identity, registration, monitoring, and policy enforcement already wired in.6 The Knowledge Lifecycle Register is a golden path for operational memory. It makes the safer choice the faster choice.
Give the Intake Registrar a literal prompt
A poetic prompt is lovely until it starts inventing retention policy. Use a fixed instruction instead.
ROLE
You are the Knowledge Intake Registrar.
TASK
Create a lifecycle record only from a linked, approved artifact or
an approved pointer to a system-of-record item.
REQUIRED FIELDS
artifact_id
artifact_type
authority_source
source_url_or_record_id
source_version
owner
approved_use
sensitivity
effective_at
review_at
expires_at
correction_route
retention_rule
hold_status
evidence_pointer
RULES
Never infer a missing owner, authority, approved use, or retention rule.
If any required field is missing, set status to needs_attention.
Do not make the item available for outward-facing use until the named
owner has approved the record.
Do not copy more sensitive content into the register than the workflow
requires.
OUTPUT
Return the completed record or a needs_attention record that names the
missing fields and the human owner required to resolve them.
Oppy's public function guide documents Public and Admin permissions plus a needs_attention workflow that can alert admins and hold outgoing chat, email, and SMS until a human responds.7 Use those controls where configured. The prompt describes good behavior. Permissions and holds create the brake.
Add a pre-use freshness check
A nightly review queue is useful. A pre-use check is better.
Before an Oppy relies on a registered item for an external response, task, or recommendation, the Freshness Checker should ask:
- Is this exact version still available at the registered source?
- Has its review or expiration date passed?
- Does a newer approved version exist?
- Do two approved sources conflict?
- Is this use permitted for this channel and purpose?
- Is the responsible owner still active?
If any answer is unknown, the system should stop the consequential step and create a named exception. A stale-answer system should fail boringly. Drama is expensive.
Oppy's recent language-access playbook states the boundary well: “Let AI coordinate the request, source, deadline, and handoff. Let qualified and authorized people own the meaning.”8 The same line should govern memory.
Do not confuse remembered context with permission to contact
A transcript may contain a phone number. A CRM note may say a customer was interested last spring. Neither fact automatically creates permission for a new campaign.
For calls or texts, run a separate communication gate immediately before contact. It should evaluate the current purpose, consent or exemption basis, do-not-call and opt-out state, applicable state rules, approved campaign, and message content. Under 47 CFR 64.1200(a)(10)-(12), the Federal Communications Commission's rules for covered robocalls and robotexts identify several ordinary opt-out terms, require reasonable revocation requests to be honored within no more than 10 business days, and permit only one non-promotional confirmation text after revocation, subject to stated conditions.9
The lifecycle register should therefore store what the artifact may support, not merely what it contains. Historical context is evidence. It is not an evergreen permission slip.
Test it on one bounded collection
Do not begin with every document your company has ever touched. That is not a pilot. It is an archaeological dig with a login screen.
Choose one collection with visible operational value and a clear human owner:
| Business | Sensible first collection | Expiration event |
|---|---|---|
| Property management | Resident communication templates and approved fee schedules | Policy revision, jurisdiction change, or owner review date |
| Mortgage | Product overlays and document-request scripts | Investor update, effective date, or compliance release |
| Title | Approved search and closing checklists | Underwriting bulletin or procedure revision |
| Brokerage | Listing intake scripts and marketing rules | MLS rule change or brokerage policy release |
| Insurance | Approved service scripts and carrier guidance | Carrier bulletin, state change, or product withdrawal |
| Dental | Membership plan terms and scheduling policies | Plan revision, fee change, or clinical owner review |
| Legal | Administrative intake instructions | Court-rule change, form revision, or attorney review |
Run the system in shadow mode for two weeks. Let it flag records without changing production behavior. Review every hold. Then enable blocking for one low-volume workflow with a documented human backup.
Measure whether memory stays trustworthy
Do not count documents registered. Count whether the system catches dangerous ambiguity before a customer does.
| Metric | What it reveals |
|---|---|
| Expired-source interception rate | How often stale knowledge was stopped before external use |
| Unknown-owner rate | How much operational knowledge lacks accountability |
| Review completion time | Whether owners can keep the register current without creating a new bureaucracy |
| Correction propagation time | How quickly an approved correction reaches affected drafts, tasks, and workflows |
| Conflict-to-handoff time | Whether uncertainty reaches the right person before a deadline |
| False-hold rate | Whether controls are precise enough for real work |
| Evidence coverage | Whether every renewal, correction, and disposition has a source pointer and approver |
A zero-hold system is not necessarily excellent. It may simply be very polite about being wrong.
The shorter letter
AI employees are becoming better at using business knowledge. The next advantage is knowing when not to use it.
Give each artifact an owner. Preserve the source. Define the approved use. Add a review clock. Separate correction from deletion. Put a human at the disposition gate. Keep communication consent in its own current control.
Then let the Oppies do what machines do well: watch dates, compare versions, trace dependencies, open cases, and refuse to turn missing authority into fluent fiction.
Memory without expiration is not institutional knowledge. It is a junk drawer with excellent search.