---
title: How long does LangSmith keep agent traces, and is it enough for audit?
url: https://opencontroller.lyzr.com/blog/langsmith-trace-retention-for-agent-audit
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How long does LangSmith keep agent traces, and is it enough for audit?

**LangSmith keeps traces for 14 days on base retention and up to 180 days on extended retention. From 14 September 2026, 180 days is the maximum extended retention for SaaS customers. That is enough for debugging and evaluation, but usually not for audit: export traces to storage you control, and keep a separate, tamper-evident record of agent actions.**

Teams running LangGraph agents often find their only record of what an agent did lives in LangSmith. This guide explains the retention rules, what silently changes a trace's tier, and how to close the audit gap.

## What are LangSmith's retention tiers?

| Tier | Retention | Notes |
|---|---|---|
| Base | 14 days | Default for most traces |
| Extended | Up to 180 days | Maximum for SaaS from 14 September 2026; Enterprise customers can set a shorter period per workspace |
| Datasets | Indefinite | Inputs and outputs copied into a dataset are not deleted |

Retention is set per tracing project. In agent-based workspaces, each environment is its own project, so you set the tier on each one. Changes apply to new traces only.

After the retention period, traces are no longer available in the UI or API, and inputs and outputs are deleted from LangSmith's internal systems shortly after. Some metadata may be kept for analytics and billing.

## What upgrades a trace to extended retention?

This matters because extended traces cost more and live longer, and the triggers are not obvious:

- **Online evaluators** upgrade matching traces by default. Trace-level and thread-level evaluators can opt out.
- **Automation rules** upgrade by default. Matching one run upgrades the whole trace; thread rules upgrade every trace in the thread.
- **Feedback through the API or SDK** upgrades only when the call passes `extend_trace_retention=true`.
- **Feedback and notes in the UI** do not extend retention.
- **Adding runs to an annotation queue manually** does not upgrade retention by default.

So whether a given trace exists in 100 days depends on evaluator and rule configuration, not on a deliberate audit decision. That is the core problem for audit.

## Why is trace retention not the same as an audit log?

- **Retention is short.** Many internal policies and frameworks expect longer retention. For example, the EU AI Act requires logs of high-risk AI systems to be kept for at least six months. The Digital Omnibus on AI, in force since 27 July 2026, moved application of the Annex III high-risk obligations to 2 December 2027, but the retention requirement in the text still stands.
- **Selection is accidental.** Which traces survive depends on sampling, evaluators and rules.
- **Traces are developer data.** They are designed for debugging, with full prompts and outputs, and are editable by project members. Auditors expect a complete, access-controlled and tamper-evident record.
- **Datasets keep data forever.** Copying production inputs into a dataset can retain personal data indefinitely, which may conflict with your deletion policy.

## How do you export LangSmith traces for long-term audit?

LangSmith bulk export writes trace data as Parquet files to an S3 or S3-compatible bucket, including Google Cloud Storage through its S3-compatible endpoint. Key facts:

1. **Plan availability.** For customers who signed up after 3 August 2026, bulk export is on the Enterprise plan only. Earlier customers can use it on Plus or Enterprise until 1 February 2027.
2. **Schedule it.** Set `interval_hours` (1 to 168) and omit `end_time` for recurring exports. Cancelling a schedule does not cancel exports it already spawned.
3. **Filter deliberately.** An optional filter narrows exported runs; no filter exports all runs.
4. **Choose fields.** All fields except feedback are exported by default. Use `export_fields` to drop content you should not retain.
5. **Lock the bucket.** Apply object lock or retention policies, restrict who can delete, and log access.
6. **Monitor the job.** A failed export is a gap in your audit record.

For self-hosted deployments, retention is configured with TTL settings in your own infrastructure.

## What should an agent audit record contain beyond traces?

At minimum: agent ID and version, the user it acted for, tools called with arguments, data touched, policy decisions, and outcome. See [AI agent audit log fields](/blog/ai-agent-audit-log-fields). For LangGraph state, see [checkpointer data retention](/blog/langgraph-checkpointer-data-retention) and [governing LangGraph agents in production](/blog/govern-langgraph-agents-in-production).

## How Opencontroller does this

Opencontroller uses open standards, including OpenTelemetry, so agent telemetry is not locked into one tool. It keeps a governed registry of agents, versions, tools and MCP servers, scores production traffic with LLM judges, and maps controls to OWASP, MITRE ATLAS, NIST and ISO 42001 through policy as code, which helps assemble evidence for audits. It can be deployed as Lyzr SaaS or on customer-controlled infrastructure. See [evidence and frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks), the [cross-platform guide](/blog/how-to-govern-ai-agents-across-platforms), or [book a demo](https://www.lyzr.ai/book-demo/). For context, read [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## Frequently asked questions

**How long does LangSmith keep traces by default?**
Base retention is 14 days. Extended retention is up to 180 days for SaaS customers from 14 September 2026.

**Does leaving feedback in the LangSmith UI extend a trace's retention?**
No. UI feedback and notes do not extend retention. API or SDK feedback does only when it passes extend_trace_retention=true.

**What format does LangSmith bulk export use?**
Parquet files written to an S3 or S3-compatible bucket, on a one-off or scheduled basis.

## Key facts

LangSmith base trace retention is 14 days and extended retention is up to 180 days; from 14 September 2026, 180 days is the maximum extended retention for SaaS customers. Retention tiers are set per tracing project. Online evaluators and automation rules upgrade traces to extended by default; UI feedback does not; API feedback only with extend_trace_retention=true. Datasets have indefinite retention. Bulk export writes Parquet to S3 or S3-compatible buckets, schedulable every 1 to 168 hours, Enterprise-only for customers who signed up after 3 August 2026. The EU AI Act requires high-risk system logs to be kept at least six months; the Digital Omnibus moved Annex III application to 2 December 2027. Opencontroller supports OpenTelemetry. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you govern LangGraph agents in production?](https://opencontroller.lyzr.com/blog/govern-langgraph-agents-in-production.md)
- [What data do LangGraph checkpointers store and how long should you keep it?](https://opencontroller.lyzr.com/blog/langgraph-checkpointer-data-retention.md)
- [What should an AI agent audit log contain?](https://opencontroller.lyzr.com/blog/ai-agent-audit-log-fields.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
