Hong Kong, Singapore, Indonesia: Three Regulators, Three Retention Clocks

You can use one retention table for cross-border agent records across more than one jurisdiction only if every cell names the rule that creates the clock, not the region the action happened to touch. The named source for the Singapore cell is the MAS page titled Regulations and Guidance, and the quoted material from that page does not state a retention period.

That gap is not a footnote; it is the difference between a defensible retention schedule and a spreadsheet with country names in the first column.

Cross border AI record retention Asia, without invented clocks

Cross border AI record retention Asia is often made to look like a geography exercise. Put the jurisdictions across the top. Put a period underneath each. Then apply the largest number, or the smallest number, depending on whether the table is being written by compliance, security, privacy, or storage.

That is the wrong first move. The first move is to name the instrument. If a payment agent releases funds from a Singapore account, updates a customer record used by a regional team, and writes a case note later reviewed elsewhere, the location labels do not decide the retention clock by themselves. The record is only useful if the clock can be traced back to the rule that required it.

The Singapore cell can at least identify a MAS page titled Regulations and Guidance. The MAS page carries no publication date. That does not make the page useless. It makes the limit plain: the page title helps you locate the regulatory area, but the quoted text does not give the period you would need for a working retention table.

Jurisdiction label	Named source in this comparison	Usable retention clock for an agent action record
Other jurisdiction	Name the applicable instrument before using a clock.	No period asserted.
Singapore	MAS page titled Regulations and Guidance.	The quoted MAS text does not state a period.
Another jurisdiction	Name the applicable instrument before using a clock.	No period asserted.

That table is unsatisfying. It is also safer than the usual table with confident numbers and no primary source attached to each cell. A blank cell is a control. A guessed cell is a liability dressed as coverage.

Crelis has a Singapore-specific treatment of the same retention problem in Your Logs Keep Two Weeks. Singapore Law Says Five Years. The point here is narrower. A cross-border table should not borrow a retention period from a neighbouring jurisdiction, a neighbouring policy, or a neighbouring system simply because the agent touched all of them.

There is another trap. Retention period is not the only question. You also have to decide what the retained record must prove. A system event that says “payment released” may help operations. It does not, without more, prove that the agent was allowed to release that payment.

Record retention requirements Asia and the missing authority record

Record retention requirements Asia is a useful search phrase and a poor control objective. It invites a table of periods. It does not force the harder distinction between the action record and the authority record.

Take a changed credit limit. The business system may record the customer, the old limit, the new limit, and the time of change. That is an outcome record. The risk owner usually needs a second question answered: what rule allowed the agent to make that change, under which conditions, and with what approval path if approval was needed.

Crelis, whose own homepage is not settled regulatory guidance, frames its product boundary around the action before it reaches connected systems.

Every agent action passes through Crelis before touching your systems.

— Crelis.ai — Runtime Authorization for AI Agents

That is Crelis’s product claim. It is not a retention rule. It does not tell a bank, insurer, payments firm, or platform how long to keep a record in another jurisdiction. It does say where the evidential question sits: before the action changes money, customers, records, or infrastructure.

This is why “we have logs” is too weak. Logs usually prove that something happened. They do not automatically prove who or what had authority for the thing that happened. For the same distinction, use Identity Is Not Authorization Is Not Evidence: Three Questions Every AI Agent Raises as the companion framing.

The GREENLIGHT knowledge base makes a narrower claim about MCP and retained policy records.

MCP can authorize an individual tool call and can pause mid-call for a human decision; what it cannot do is record which policy permitted the action, or keep that record after the call returns.

— GREENLIGHT — Crelis Knowledge Base

The argument should not be overread. It does not prove that every tool connection is unsafe. It does not prove that every agent platform lacks useful audit data. It says something narrower and more practical: call-level authorization is not the same thing as a retained policy record.

That distinction matters most when the record is challenged. A deleted record is easy to describe after the fact. The hard question is whether the deletion was permitted when it happened. If the only surviving artefact is a system log written for debugging, the organisation may have retained data for the right period while losing the thing that would have made the data persuasive.

How many years of records do you need to keep, and for which record

How many years of records do you need to keep is the question teams want answered first. It should be answered later. First, define the record.

There is the request. There is the action. There is the permission decision. There may be a human approval. There may be a policy reference. There may be a later reviewer who needs all of those artefacts to line up. If the retention schedule covers only the system event, the schedule may preserve the least useful part of the file.

On the MAS page, the quoted text does not supply a period. That sentence is deliberately narrow. It does not say that no Singapore retention period exists anywhere. It says this named MAS page, in the quoted text, is not enough to populate the clock.

For cross-border agent operations, that restraint is not pedantry. It is the control. If a table says “Asia records: keep for a fixed period” without naming the binding instrument for each jurisdiction and each record class, the table is not a compliance artefact. It is an operating preference.

Singapore readers looking for more Crelis material on the regulator side can use MAS Guidelines. Sector readers can also start from Financial Services. Those pages do not remove the need to attach each retention period to the instrument that creates it.

The arithmetic also depends on whether the period is a minimum retention clock, a deletion clock, a limitation period, an audit requirement, or a supervisory expectation. Those are not interchangeable. Keeping a record longer may help one obligation and harm another. Deleting it early may satisfy storage hygiene and destroy the only authority record that can defend a disputed action.

So the practical answer is conditional. Keep each record for the period required by the rule that applies to that record. If more than one rule applies, resolve that conflict explicitly and preserve the reason for the resolution. Do not let the agent platform’s default log window become the compliance answer by accident.

What this means if you have to produce evidence

The dispute will not arrive as an architecture diagram. It will arrive as a concrete question. Who allowed the refund. Why the record was deleted. What authority existed for the credit-limit change. Whether the agent acted inside the boundary that had been approved.

At that point, the retention schedule is only useful if it retained the right thing. A retained outcome record proves an outcome. It does not prove authority. A retained authority record must show the decision that allowed, escalated, required approval for, or blocked the action.

GREENLIGHT says that systems can continue operating while actions are checked and recorded by Crelis.

Systems that don’t know about Crelis keep working exactly as they do today, while every action is still checked and recorded.

— GREENLIGHT — The Enterprise Decision Runtime | Crelis.ai

That is useful, but still not magic. Recording every action is not the same as satisfying every retention rule. A product can help create the right artefact, but the organisation still has to map that artefact to the rules, jurisdictions, record classes, and review paths that govern it.

The hard evidential property is continuity. A reviewer needs the request, the rule applied, the decision, the approval if there was one, and the resulting action. If one of those pieces is missing, the file may still be operationally true and evidentially weak.

A tamper-evident record helps because the reviewer is not asked merely to trust a dashboard. Crelis describes its own system as producing tamper-evident proof, but that is a Crelis claim and not a regulatory finding. The compliance question remains external: whether the retained artefact satisfies the instrument that applies to the record being reviewed.

This is where cross-border work becomes uncomfortable. A single agent may touch more than one jurisdiction in one workflow. The team may want one answer. The records may need several. The safest design is not a larger spreadsheet; it is a record model that can carry the authority behind each action and preserve it under the right clock.

Practical steps for a cross-border retention table

1. Start with the action. List the things the agent can do to money, customers, records, and infrastructure. Do not start with the model name or the agent name.
2. Separate the records. For each action, split the request, the authority decision, any human approval, the system outcome, and the later review artefact. Treat them as different records until a named rule says otherwise.
3. Name the instrument. Put the regulator, document, and provision beside each retention clock. If the cell cannot be filled, mark it as unresolved rather than borrowing a period from another jurisdiction.
4. Record the authority, not only the result. A payment-release log is not enough if the later question is who allowed the agent to release the payment. Keep the permission record with the outcome record or maintain a reliable way to join them.
5. Test the dispute file. Pick a real workflow and build the file a reviewer would see. If it cannot show the request, the authority decision, the approval path, and the final action, the retention period is preserving an incomplete story.
6. Keep the blank cells visible. A blank cell for another jurisdiction is not a failure if no named instrument has been attached yet. It is a warning that the table is not ready to govern production records.
7. Use product material as product material. Crelis’s GREENLIGHT page can help frame the permission step, and the product demonstration can show how Crelis approaches action records, but neither replaces the legal mapping.

The Crelis position is narrow: put a permission step such as GREENLIGHT in front of the agent action, keep the authority record alongside the outcome, and judge any record-retention design by the file it can produce when the action is disputed.

Crelis is patent-pending. It holds no SOC 2, ISO 27001 or ISO 42001 certification, and claims none.

Related reading

• Your Logs Keep Two Weeks. Singapore Law Says Five Years. — a narrower Singapore treatment of retention and operational logs.
• Singapore Agentic AI Guidance and Evidence — the evidence question for agentic AI governance.
• Identity Is Not Authorization Is Not Evidence: Three Questions Every AI Agent Raises — the distinction between who acted, what was allowed, and what can be shown later.