VibeMemBench

Evaluating memory systems for coding agents on real repository coding tasks.

• Code: https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/VibeMemBench
• Paper: https://arxiv.org/abs/2609.23570

Coding agents work on real repository tasks, and persistent memory systems promise to reuse experience across them. VibeMemBench tests whether that experience actually improves executable repository work. It pairs two runs of the same task under one fixed agent, tools, sandbox, and budget. Only the memory condition changes. Executable tests decide whether the task is resolved.

Overview. Top: offline construction of the frozen verified experience. Bottom: the two released evaluation tracks.

The dataset has two parts:

Part	File	Size	Role
Targets	test.parquet	1 MB	111 coding tasks to solve, in SWE-bench format
History	history/	2.2 GB	3,634 completed agent trajectories from the same repositories, the memory substrate

The targets span 90 repositories and 9 languages. The history is what a memory system reads to build experience. A target is never allowed to see history from its own future, see Causal rule.

Download

load_dataset gives you the targets only. The history is not a dataset split, it is a directory of files, so it needs snapshot_download.

Targets only (1 MB, enough to inspect the tasks):

from datasets import load_dataset

targets = load_dataset("Tongyi-ConvAI/VibeMemBench", split="test")
print(targets[0]["instance_id"], targets[0]["problem_statement"][:200])

Everything (2.2 GB, needed to run the benchmark):

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Tongyi-ConvAI/VibeMemBench",
    repo_type="dataset",
    local_dir="benchmark",
)

One language only, if you do not want all 2.2 GB:

snapshot_download(
    repo_id="Tongyi-ConvAI/VibeMemBench",
    repo_type="dataset",
    local_dir="benchmark",
    allow_patterns=["test.parquet", "history/python/*"],
)

Or from the command line:

hf download Tongyi-ConvAI/VibeMemBench --repo-type dataset --local-dir benchmark

The harness expects the result at benchmark/, so local_dir="benchmark" puts files exactly where the code looks for them.

Layout

test.parquet                                      # 111 targets
history/<language>/<repo>/<instance_id>/<instance_id>.traj.json

Example history path:

history/python/adamchainz__flake8-comprehensions/adamchainz__flake8-comprehensions-143/adamchainz__flake8-comprehensions-143.traj.json

<repo> is the upstream owner/name with / replaced by __.

Counts

Language	Targets	History trajectories
Python	56	2043
JavaScript	16	503
Go	12	263
Rust	10	185
TypeScript	5	383
Kotlin	4	170
Swift	4	47
Java	2	34
PHP	2	6
Total	111	3634

Language (left) and primary task type (right) distribution of the 111 targets.

Targets cover bug fixes, feature requests, interface changes, and configuration work. Bug fixes account for under a third of them. Upstream pull requests range from 2016-03 to 2025-09.

Targets: test.parquet

One row per target, 16 columns, SWE-bench format. The Visibility column states what the evaluation protocol allows the solver to see.

Column	Type	Visibility	Meaning
instance_id	str	solver	Unique id, ends in the upstream PR number
repo	str	solver	Upstream owner/name
base_commit	str	solver	Commit the task starts from
problem_statement	str	solver	The issue text the agent must solve
image_name	str	solver	Docker image with the prepared environment
language	str	solver	One of the 9 languages
install_config	dict	solver	Build and test recipe, see below
pr_description	str	hidden	Upstream PR text, leaks the intended fix
created_at	str	hidden	Upstream PR timestamp
interface	str	hidden	Interface change summary
license	str	hidden	Upstream repository license
meta	dict	hidden	Construction annotations
patch	str	evaluator	Gold patch, the answer
test_patch	str	evaluator	Adds the tests that decide the outcome
FAIL_TO_PASS	list	evaluator	Tests that must flip to passing
PASS_TO_PASS	list	evaluator	Tests that must stay passing

install_config carries base_image_name, docker_specs, install, test_cmd, and log_parser. The last two are what an evaluator needs: run test_cmd in the image, then parse its output with the named parser to decide which tests passed.

A run is resolved when every test in FAIL_TO_PASS and PASS_TO_PASS passes.

History: history/

Each .traj.json is one completed agent trajectory for an earlier issue in the same repository. The top level is a single key:

{"messages": [{"role": "...", "content": "...", "extra": {...}}, ...]}

Message roles are system, user, assistant, tool, and exit. Assistant messages may carry reasoning_content and tool_calls; extra may carry the raw model response, the shell actions, the command raw_output, usage, cost, and a timestamp.

Two things to know before you build on this:

• extra.timestamp is not the PR date. It is the wall clock of the machine that produced the trajectory, so it says nothing about when the upstream issue happened. Use the trailing number of instance_id for chronology.
• One trajectory is 98 MB (platers__obsidian-linter-1157), and a second is 78 MB. Load history lazily or cap the size, or a naive loop will exhaust memory.

Causal rule

A target may only be given history from earlier issues in the same repository. Chronology comes from the trailing number of instance_id, where a smaller number means earlier:

feed history H to target T  iff  repo(H) == repo(T)  and  number(H) < number(T)

This matters in practice, because the history directory also contains the target's own trajectory. For target adamchainz__flake8-comprehensions-553 the repository directory holds 143, 179, 195, 197, 200, 205, 553. Only the first six may be fed. Passing 553 to a memory system would let the agent retrieve the answer to the task it is being tested on.

Licensing

The license column records each upstream repository's license. Across the 111 targets: MIT 45, Apache-2.0 29, project-specific 20, BSD-3-Clause 10, MIT-0 5, BSD-2-Clause 1, unspecified 1. Problem statements, patches, and tests are derived from those public repositories and remain under their original terms. Check the column before redistributing any single instance.

Privacy

The trajectories were audited for leaked secrets and personal identity. They contain no real API keys, tokens, passwords, or operator identity. Values that look like credentials are public documentation examples, such as the canonical AWS example key used in test fixtures. Usernames and email addresses come from the public issues and commits of the evaluated repositories.

Citation

@article{fan2026vibemembench,
  title   = {VibeMemBench: Evaluating Memory Systems for Coding Agents on Real Repository Coding Tasks},
  author  = {Fan, Liyang and Shi, Yingcheng and Li, Yongbin and Sun, Chenghao and Chen, Xin and Xu, Xander and Wei, Hu and Ni, Shiwen and Yang, Min and Ye, Jieping},
  journal = {arXiv preprint},
  year    = {2026},
  note    = {arXiv:2609.23570}
}

VibeMemBench builds on SWE-rebench V2 for the repository tasks and uses mini-swe-agent as the fixed agent.