How long to keep AI system logs: EU AI Act, NIST and HIPAA

Six months is the EU AI Act's floor for high-risk systems. What sets your real period, and what each record should hold.

11 min read
In short: The EU AI Act sets at least six months for logs of high-risk systems, for providers (Article 19) and deployers (Article 26(6)), unless other law says otherwise. NIST AI RMF sets no period, and HIPAA keeps documentation for six years. Keep the decision your control made, with sensitive values removed before the record is written.

How long should you keep AI system logs?

The EU AI Act requires deployers of high-risk systems to keep the systems' logs for at least six months (Article 26(6)). Providers must keep them too (Article 19), and must design systems to record events automatically (Article 12). Other law can require longer: HIPAA's documentation rule is six years. Keep the decision your control made, not just the request.

Six months is a floor, and only for high-risk systems. Most AI systems a software company runs, such as coding agents and chat assistants, aren't high-risk, so the Act sets no period for them. NIST AI RMF sets no period either. Your retention period then comes from three places: the laws that apply to the data in the logs, what your customers and auditors expect, and how far back you need to look when something goes wrong.

This guide covers what each framework asks, what an AI system log should contain, how to keep it without turning it into a copy of your most sensitive data, and where to keep it. For the governance program these logs feed, see what AI governance is.

What does each framework require for AI log retention?

Only the EU AI Act sets a number specific to AI systems, and only for high-risk ones. The other frameworks ask for monitoring and records, and leave the period to you or to other law. The following table summarizes each source in our own words.

SourceWhat it asksPeriod
EU AI Act, Article 12

High-risk systems must record events automatically over their lifetime, enough to trace risks, support post-market monitoring and let deployers monitor operation

None; it's a design requirement on providers
EU AI Act, Article 19

Providers keep the automatically generated logs under their control

At least six months, unless other law says otherwise
EU AI Act, Article 26(6)Deployers keep the logs under their controlAt least six months, unless other law says otherwise
NIST AI RMF 1.0, MANAGE 4.1

Post-deployment monitoring is planned and carried out, including response, recovery and decommissioning

None set
HIPAA, 45 CFR 164.312(b)

Systems that hold or use electronic protected health information have mechanisms that record and examine activity

None set in the standard
HIPAA, 45 CFR 164.316(b)(2)Security Rule documentation is retained

Six years from creation or from when it was last in effect, whichever is later

GDPR, Article 5(1)(e)

Personal data is kept no longer than its purpose needs (storage limitation)

An upper bound, not a floor

The EU AI Act's six months comes with a qualifier that matters: other EU or national law can set a different period, and the Act names data protection law in particular. A log full of personal data can't be kept forever because the AI Act sets a minimum. The General Data Protection Regulation (GDPR) pulls the other way, so decide the period with your data protection team. Logs that hold less personal data are easier to keep longer.

The Act's duties attach only to logs under your control. A deployer of a third-party high-risk system often holds only part of the logs, and the provider holds the rest. Agree in the contract who keeps what, and how you get the provider's logs when you need them.

The EU AI Act sources are Regulation (EU) 2024/1689 and its July 2026 amendment, Regulation (EU) 2026/1744. The amendment didn't change Articles 12, 19 or 26(6), but it moved their application date to December 2, 2027 for Annex III systems and to August 2, 2028 for Annex I systems (EUR-Lex, read on November 16, 2026). For the full timeline, see the EU AI Act for deployers. The NIST source is the AI Risk Management Framework.

HIPAA's six years applies to documentation the Security Rule requires, such as policies and risk assessments. It isn't a stated period for audit logs, though many organizations apply it to them. For HIPAA in full, see HIPAA compliance for AI agents.

What should an AI system log contain?

An AI system log should record who asked, which system acted, what it tried to do, and what your control decided and why. A request log alone shows that something happened. The decision record shows that a control evaluated it, which is what an auditor or an incident review needs.

The following table lists the fields that make a record useful as evidence.

FieldWhy it matters
Time

Orders events and places them inside or outside a retention window

AI system and actorTies the event to an entry in the inventory and to its owner
PersonIdentifies whose request or session the action ran under
Action and targetNames the tool, model provider or resource, as a stable label
DecisionAllow, deny or hold, with the rule that fired and its reason
Mode

Whether the rule ran in dry run, so pre-enforcement results are visible

Policy versionWhich version of the rule decided, and who published it
Correlation identifier

Joins the steps of one task, and joins your records to the control's

Sensitive-information findingsThe types of data found, not the values

Article 12 asks high-risk systems to log enough to identify situations that could present a risk, to support post-market monitoring, and to let deployers monitor operation. A record with the decision, its reason and the person behind the request covers all three better than raw prompts and responses do. For a field-by-field design, see how to audit and log AI agent activity.

How do you keep AI logs without copying sensitive data into them?

Keep the type of sensitive data and drop the value before the record is written. A log that stores every prompt and tool argument verbatim becomes a second copy of your customer data, with wider read access and a longer retention period than the system it came from.

The following practices keep a log useful without making it a liability:

  • Detect before you store. Run sensitive-information detection on prompts, responses and tool arguments, and write the entity types and their positions instead of the matched values.
  • Use server-controlled metadata. Attach identifiers, enumerated types, status codes and amounts. Don't attach free text that a user or a model wrote.
  • Label actions with stable names. A label such as tools.issue-refund is searchable for years. A copy of the arguments isn't needed to show that a refund was denied.
  • Separate the decision from the payload. If you must keep raw payloads for debugging, keep them on a short period in a separate store, and keep the decision record on the long one.

This also resolves most of the tension with data protection law. A decision record that holds identifiers and entity types, rather than personal data in the clear, is easier to justify keeping for six months or six years.

Should you keep AI system logs in Arcjet or in your SIEM?

Keep them where your retention obligation can be met and shown, which for many teams is both. Arcjet holds the operational view and can hold the long-term copy. Your security information and event management (SIEM) system, or your own storage bucket, holds the copy under your own retention policy.

The following table compares the options.

OptionRetentionBest for
Arcjet Console decision historyFollows your plan's log retentionDay-to-day review, investigation and dry-run analysis
Arcjet long-term retentionOpt-in, beyond the plan's window

Teams with a log-retention duty and no SIEM, or that want the record next to the inventory

SIEM export (Enterprise)Your SIEM's policy

Detection rules, alerting and a long archive in your SIEM or log store through a log drain

Whichever you choose, write the decision down: the period, the store, and the obligation it meets.

What Arcjet does not cover

Arcjet keeps the decisions it makes and the records it receives. It doesn't keep every log your AI systems produce, and it doesn't decide your retention period. The following stays with you:

  • Choosing the period. The legal analysis that weighs the AI Act's minimum against data protection law and sector rules.
  • Third-party system logs. A high-risk system that runs in a vendor's cloud generates its own logs. Arcjet doesn't hold them.
  • Application logs. What your agent did outside the calls Arcjet decides, such as its internal reasoning or the results of allowed tool calls.
  • Systems Arcjet can't see. Records exist only for enrolled macOS and Windows devices, applications and agents that report to Arcjet, and coding agents with hooks installed.
  • Your SIEM's retention. Once you export a record, its period is set by your SIEM or bucket policy.

Where Arcjet fits

Arcjet records the decision your control made on each action, which is the part of an AI system log that most teams lack. Guard decides tool calls, prompts and model use, and each decision carries its conclusion, the rule, the reason, the action's label, your metadata and an optional correlation identifier. Dry-run decisions are recorded too, and every policy change names the person who made it.

Sensitive-information detection covers PII, secrets and PHI. When the server-side detector flags a value, Arcjet removes it before storing the decision and keeps the entity types and their positions. The record shows that a call carrying a medical record number was denied without containing the number.

Each record belongs to an AI system in the AI inventory, with its owner and purpose. Console decision history follows your plan's log retention. Opt-in long-term retention keeps decisions for longer for teams with a retention duty. On Enterprise, a log drain also exports decisions to your SIEM or log store.

The evidence page maps retention and runtime records to EU AI Act Articles 19 and 26(6) and to NIST MANAGE 4.1, alongside the rest of the framework. To work through the AI Act's obligations one by one, use the EU AI Act checklist. For how the evidence reaches your GRC tool, see AI governance evidence export.

Frequently asked questions

Does the six-month log rule apply to coding agents and chat assistants?

Usually not. Articles 19 and 26(6) apply only to high-risk AI systems, and coding agents and chat assistants used to write code or text aren't high-risk under Annex III. Your retention period for them comes from other law, customer contracts and how far back your incident reviews need to look.

Can GDPR stop me keeping AI logs for six months?

It can limit what you keep. The AI Act's six months applies unless other law says otherwise and names data protection law in particular, and GDPR's storage limitation principle bars keeping personal data longer than needed. Records that hold identifiers and entity types instead of personal data in the clear are easier to keep.

Does HIPAA require six years of audit logs?

HIPAA requires six years of retention for Security Rule documentation, such as policies and risk assessments, under 45 CFR 164.316(b)(2). The audit controls standard, 45 CFR 164.312(b), sets no period for activity logs, though many organizations apply six years to them too. Decide the period with your compliance team.

What is the difference between a request log and a decision log?

A request log shows that something happened. A decision log shows what a control decided about it, with the rule and the reason, which is the evidence an auditor or incident review needs. Keep both joined by a correlation identifier.

Who keeps the logs of a third-party high-risk AI system?

Both parties keep the logs under their own control. The provider keeps its share under Article 19 and the deployer its share under Article 26(6), each for at least six months. Agree in the contract who keeps what and how the deployer gets the provider's logs when needed.

AI runtime security in your code

Protect your AI agent workflows with Arcjet

A retention period only helps if the record shows what your control decided. Record guard decisions, with sensitive values removed, alongside what the application did next.