---
title: "Benchmarks | Total Recall"
description: "Total Recall benchmark results: 98.0% end-to-end on LongMemEval (500 questions, gpt-5.4), 97.73% Recall@10 retrieval-only, and 92.36% Recall@10 on LoCoMo. Full methodology, public comparison context, and reproducibility."
canonical: "https://total-recall.dev/benchmarks"
last-updated: "2026-09-12"
---

# Benchmarks

Three questions, three numbers. Total Recall is measured on public long-memory benchmarks.

## At a glance

- **LongMemEval, end-to-end:** 98.0% judged answer accuracy (500 questions, gpt-5.4, measured on an earlier retrieval snapshot; rerun pending).
- **Retrieval-only Recall@10:** 97.73% (deterministic local retrieval).
- **LoCoMo Recall@10:** 92.36% (retrieval-only, v1.9.81).
- **MEME, current fact above outdated:** 99 of 100, with the right memory in the top 5 for 692 of 692 questions (retrieval-only).

## LongMemEval

LongMemEval tests whether a system can answer questions that depend on information spread across a long history of sessions. The end-to-end number reflects judged answer accuracy; the retrieval number reflects whether the correct evidence was surfaced in the top 10 results.

## LoCoMo

LoCoMo is a long-conversation memory benchmark. The reported number is retrieval Recall@10, measured on v1.9.81.

## What these numbers do and do not mean

They measure retrieval quality and end-to-end answer accuracy on public benchmarks. They are not a claim about every possible workload. Where a number comes from an older snapshot or a specific model, that is stated inline.

## Reproducibility

Methodology and benchmark harness details are published so results can be reproduced. See the full benchmarks page at https://total-recall.dev/benchmarks.