The highest-scoring agent memory on four independent benchmarks.

Agent memory is how an AI agent keeps your rules, facts and past work, and uses them on the next task. An agent that forgets or mixes them up makes the wrong call. We test Trail on public benchmarks against the best published results, with the same questions, the same AI model and the same scoring.

How to read these results

  1. What a benchmark is

    A public test built by researchers. Every system answers the same questions and is marked the same way, so scores can be compared.

  2. What we compare

    Trail against the best published result on each test, using the same AI model. The model is the same on both sides, so the difference comes from memory.

  3. What a point means

    Scores are out of 100. A 13-point lead means Trail got about 13 more tasks right out of every 100.

The results

  • 01From Microsoft

    STATE-Bench

    Does the agent follow business procedures correctly?

    150 real tasks across travel booking, customer support and shopping. The agent learns each procedure from past examples, then has to carry it out.

    +13.2 points ahead of the next best
    Trail
    67.7%
    Microsoft Foundry Memory
    54.5%

    Both sides used GPT-5.4

  • 02#1 on conflict resolution

    MemoryAgentBench

    When rules and facts change or conflict, does the agent use the right one?

    Tests recall, conflict resolution and learning on the job across 14 datasets. Trail ranks first on the conflict resolution track. The score shown is the overall average, out of 100.

    +8.2 points ahead of the next best
    Trail
    68.8
    The benchmark authors' best result
    60.6

    Both sides used GPT-5-mini

  • 03

    LongMemEval

    Does the agent remember and update what it learned over time?

    500 questions about details spread across many past sessions, including facts that changed along the way.

    +9.2 points ahead of the next best
    Trail
    95.2%
    Oracle Agent Memory
    86.0%

    Both sides used GPT-5.5

  • 04

    LoCoMo*

    Can the agent recall details from very long conversations?

    1,540 questions about conversations that run across weeks of back-and-forth.

    +11.6 points ahead of the next best
    Trail
    90.7%
    Zep
    79.1%

    Both sides used GPT-4.1-mini

How we count a result

Scores move as systems improve, so we only claim a result that passes every check below.

  • The official test, run in full. Every question answered, none skipped.
  • The same AI model and the same marking as the published result we compare against.
  • Code pinned to a tracked version, so every run can be repeated.
  • Runs that miss any of these aren't counted.

Mem0 reports a higher LoCoMo score, but it was marked by a more lenient grader. A different grader makes it a different test, so we don't compare against it.

We've submitted these results to the benchmark maintainers. They haven't been published yet, so Trail won't appear on the public leaderboards for now.

Results as of November 2026.

Benchmarks are one test.
Your rules are the real one.

Bring a process where rules conflict and we'll show you which one Trail applies, and why.