LoCoMo (Mirror)

Mirror of the LoCoMo benchmark (CC BY-NC 4.0), converted to parquet for direct load_dataset use. Content is identical to the original.

Dataset Overview

10 multi-session conversations (two speakers, up to 32 sessions, ~5,900 turns with timestamps), annotated with 1,986 QA pairs across five categories of memory ability:

Category	Count	Ability Tested
1 single-hop	282	Single-hop fact recall
2 temporal	321	Temporal reasoning
3 multi-hop	96	Multi-hop reasoning across sessions
4 open-ended	841	Open-ended synthesis
5 adversarial	446	Adversarial (questions about events that never happened)

Data Structure

Each row is one conversation sample:

Field	Type	Description
sample_id	string	Conversation id, e.g. conv-26
qa	string (JSON)	List of QA pairs
conversation	string (JSON)	Structured conversation
event_summary	string (JSON)	Per-session event summaries
observation	string (JSON)	Per-session observations
session_summary	string (JSON)	Per-session summaries

The structured fields are stored as JSON strings (byte-equivalent to the original file); decode them with json.loads after loading.

Note: category-5 adversarial questions have no answer field — their ground truth lives in adversarial_answer.

Usage

import json
from datasets import load_dataset

ds = load_dataset("CodyTonn/LoCoMo", split="train")

sample = ds[0]
conversation = json.loads(sample["conversation"])
qa_pairs = json.loads(sample["qa"])

for q in qa_pairs:
    answer = q.get("answer", q.get("adversarial_answer"))  # adversarial questions fall back to adversarial_answer

Citation

@article{maharana2024evaluating,
  title   = {Evaluating Very Long-Term Conversational Memory of LLM Agents},
  author  = {Maharana, Ady and Lee, Dong-Ho and Turek, Sergey and Bansal, Mohit},
  journal = {arXiv preprint arXiv:2402.17753},
  year    = {2024}
}

• Original paper: arXiv:2402.17753