LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
Paper • 2410.10813 • Published • 15
Datasets:
The Dataset Viewer has been disabled on this dataset.
Blog Experiments
Data files that sit behind the technical blog posts of Minh Le Duc (Lê Đức Minh). Each post that runs an experiment gets a folder here, so a reader can rerun the numbers instead of trusting them.
The blog covers LLM agents, retrieval and memory, Vietnamese NLP, encoder benchmarking and small-model evaluation. New posts go to ZyVOP first; older ones are on Medium.
Contents
| Folder | File | Used by |
|---|---|---|
agent-memory-tax/ |
longmemeval_oracle.json |
Draft post "Three Papers Say Your Agent's Memory Loses to Plain RAG" (a laptop test of where agent memory loses to plain retrieval). The link will be added here once it is live. |
More folders are added as posts ship.
agent-memory-tax/longmemeval_oracle.json
The oracle split of LongMemEval (cleaned release): 500 questions, each paired only with the chat sessions that contain the evidence. It is the "easy retrieval" setting: no distractor sessions, so it isolates what a memory layer does to the evidence rather than whether it can find it.
Provenance. This is a byte-for-byte copy of longmemeval_oracle.json from xiaowu0162/longmemeval-cleaned (SHA-256 821a2034d219ab45846873dd14c14f12cfe7776e73527a483f9dac095d38620c, 15,388,478 bytes). Nothing was filtered, rewritten or re-labelled. It is mirrored here only so the blog's scripts point at one stable place.
Fields (one JSON object per question):
| Field | Meaning |
|---|---|
question_id |
Question identifier |
question_type |
One of six types (counts below) |
question / answer |
The question and its gold answer |
question_date |
When the question is asked in the simulated timeline |
haystack_sessions |
The chat sessions (lists of role / content turns) |
haystack_session_ids / haystack_dates |
Session ids and timestamps, aligned with haystack_sessions |
answer_session_ids |
Which sessions hold the evidence |
Question types: temporal-reasoning 133, multi-session 133, knowledge-update 78, single-session-user 70, single-session-assistant 56, single-session-preference 30.
import json
from huggingface_hub import hf_hub_download
path = hf_hub_download("LakoreAI/experiments", "agent-memory-tax/longmemeval_oracle.json", repo_type="dataset")
data = json.load(open(path))
print(len(data), data[0]["question_type"])
The dataset viewer is off because haystack_sessions is deeply nested; load the JSON directly.
License and attribution
MIT, inherited from the upstream release. All credit for the benchmark goes to its authors. If you use the data, cite the original paper, not this mirror:
Wu et al., LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory, 2024. https://arxiv.org/abs/2410.10813
Where the blog lives
- ZyVOP (canonical for new posts): https://zyvop.com/author/l445
- Medium (legacy): https://medium.com/@minhle_0210
- GitHub: https://github.com/MinLee0210
- LinkedIn: https://www.linkedin.com/in/minhle007/
ZyVOP (13 posts)
Medium (49 posts)
Published on medium.com/@minhle_0210 and the Artificial Intelligence in Plain English publication. Dates marked n/d were not recorded. Taken from the author's own index, which is a floor rather than a complete list.
- Downloads last month
- 8