BENCHMARK · MULTI-AGENT COLLABORATIVE MEMORY

Benchmarking
Collaborative Memory Boundaries
across Workflow Topologies

Does task-relevant memory reach the agents that need it — while inapplicable content stays isolated? CoMemBench turns source-grounded multi-agent workflows into matched clean–polluted memory experiments with typed, version-aware artifact handoffs.

800
Workflow families
4
Source-grounded domains
800
Polluted instances
6–19
Nodes per family (median 11)
01 — Overview

Not another single-agent benchmark

Agents form local memory from their own tasks, observations, and execution histories. In a workflow, whether that memory should cross an agent boundary is task-conditioned: a downstream node either requires the artifact, or must stay isolated from it. CoMemBench evaluates exactly that boundary.

①

Task-conditioned boundaries

Applicability is defined by workflow dependency edges, not by speakers, roles, or privacy policies. A document required by node A is a distractor for node B when no edge connects them.

②

Matched clean–polluted pairs

800 pollution instances are paired with controls. If inapplicable content changes downstream behavior, the matched pair exposes it — without equating benchmark-visible artifacts with internal memory.

③

Typed, version-aware handoffs

Artifacts carry types and versions (repository commits, Lean toolchains, corpus revisions). Verifiers check that what crossed a boundary is what the receiving node declared it needed.

02 — Domains

Four domains, 200 families each

Every family is a directed workflow of typed artifact handoffs, grounded in a real source benchmark and executed with domain-native verification.

TB
Toolathlon / Toolathlon-GYM

Stateful Tool Use

Long-horizon application workflows with live tools — search, finance APIs, office documents, messaging. Memory includes tool state, intermediate artifacts, and receipts that later nodes depend on.

200 familiesterminal + API toolsreceipt-gated
SW
SWE-bench Verified

Repository Code Change

Real GitHub issues turned into multi-node workflows: reproduce → localize → patch → test → release. Version-aware handoffs pin commits so a stale analysis cannot silently satisfy a downstream node.

200 familiescommit-pinnedregression verified
IR
BrowseComp+

Offline Information Retrieval

Hard multi-clue questions over a fixed candidate corpus. Agents gather evidence partitions, reconcile entities, and return answers with auditable provenance receipts — often under bounded revision loops.

200 familiesfixed corpusprovenance receipts
L4
Lean 4 / HepLean

Mathematical & Structured Reasoning

Real proof-dependency graphs from HepLean. A downstream agent may only receive a successfully compiled ProofReceipt whose theorem ID, statement hash, commit, and toolchain match its declared incoming edge.

200 familiesProofReceipt-gatedno sorry / axioms
03 — Related Work

A chronology of memory & multi-agent benchmarks

From static evidence use toward governed multi-party memory — and from interaction topologies toward executable task dependencies. CoMemBench sits at the intersection: task-conditioned memory boundaries on executable dependency workflows. Click any card to open the paper.

Tracks:
Thread A

Agent memory & information boundaries

LongBench, RULER hsieh2024ruler and KILT / BEIR established static evidence use. MemoryBank, LoCoMo, LongMemEval, RealTalk, StoryBench moved to persistent and interactive memory; MemoryAgentBench added selective forgetting. GroupMemBench, GateMem, PAC-Bench then asked who may see what.

Thread B

Multi-agent & workflow structure

AgentBench, MindAgent, LLM-Coordination study coordination; MultiAgentBench varies star / chain / tree / graph interaction protocols zhu2025multiagentbench; TaskBench and MAS-Orchestra vary task dependencies — chains, DAGs, depth, breadth, parallelism.

Thread C

Executable outcomes

τ-bench, AppWorld, Toolathlon li2025toolathlon and Toolathlon-GYM make success checkable in real environments. PeopleJoin and SILO-BENCH add distributed information. CoMemBench combines these lines with typed handoffs and matched pollution pairs.

04 — Task Examples

Four worked examples, one per domain

Verbatim excerpts from the v1 release (workflow_families.jsonl). The goal is what every agent in the family is ultimately working toward.

01
Stateful Tool Use

Evidence-based workplace nutrition program

v15_family_0001_b9f32b72e8fd 19 nodes · 60 edges composite topology Toolathlon
Global goal (excerpt)
You are a corporate wellness coordinator tasked with designing an evidence-based workplace nutrition program. Your goal is to combine practical recipe data with published nutritional science research to create a scientifically grounded meal program for employees.

Start by reading the program_requirements.md file in your workspace, which describes the wellness program goals, target population size, and budget constraints. Also read the dietary_restrictions.csv file that lists common employee die
Agent roles
retrieval_agent ×10 analysis_agent ×2 artifact_agent ×2 implementation_agent ×3 workflow_verifier
Artifact types on handoff edges
source_evidence code_artifact execution_receipt form_receipt tabular_artifact document_artifact
Notes
The deepest chain in the release: a 16-stage pipeline that starts from three parallel sources, writes and executes three Python programs, then iterates a live spreadsheet across four handoffs before a Word proposal and a validation script. Nine revision edges let a failed validation send the workflow back into the workbook trail.
Workflow topology
program brief recipes DB nutrition research requirements.md categorize_recipes.py run categorize analyze_research.py run analysis build_menus.py run build_menus preferences form analysis.xlsx update workbook update workbook refine menus finalize workbook proposal.docx validate_menus.py workflow verifier
handoff flow revision (bounded re-entry)
02
Repository Code Change

Sphinx literalinclude indent repair

v15_family_0211_750658a4e4ee 16 nodes · 30 edges sphinx-doc__sphinx-10323 SWE-bench Verified
Global goal (excerpt)
Resolve the following real issue in repository sphinx-doc/sphinx at base commit 31eba1a76dd485dc633cae48227b46879eda5df4. Work in the fixed repository environment. Reproduce the reported behavior, localize its cause, implement the smallest correct patch, run relevant regression tests, and return a release decision.

Use of literalinclude prepend results in incorrect indent formatting for code eamples ### Describe the bug

Cannot determine a mechanism to use literalinclude directive with `prepe
Agent roles
issue_analyst reproduction_agent ×3 repository_analyst ×3 regression_agent ×2 patch_designer implementation_agent test_agent ×3 release_agent workflow_verifier
Artifact types on handoff edges
issue_requirements reproduction_receipt localization_receipt baseline_receipt patch_plan patch_candidate test_receipt release_bundle
Notes
A wide split-join in the middle: three independent reproductions, three independent localizations, and a pre-change baseline all feed one patch plan. Four verification receipts (three focused tests plus a regression rerun) gate the release bundle before the trusted SWE-Bench container accepts it.
Workflow topology
issue analyst reproduce 1 reproduce 2 reproduce 3 localize 1 localize 2 localize 3 baseline patch plan implement patch test 1 test 2 test 3 regression rerun release bundle workflow verifier
handoff flow revision (bounded re-entry)
03
Offline Information Retrieval

Twelve-partition evidence reconciliation

v15_family_0413_cef4321720b5 17 nodes · 41 edges browsecomp_plus/test/22 bounded_revision_loop
Global goal (excerpt)
Answer this real information-seeking question from the released fixed candidate corpus. Use exact supporting passages with document identities, reconcile all clues to one candidate, and return the requested attribute with an auditable provenance receipt.

There is a sport in the world where teams from a certain continent have dominated the biggest yearly international competition in the sport until 2022, excluding one event. The first of these official biggest yearly international competitions f
Agent roles
evidence_agent ×12 entity_resolution_agent answer_agent provenance_auditor result_synthesizer workflow_verifier
Artifact types on handoff edges
source_evidence resolved_candidate answer_candidate provenance_receipt retrieval_result_bundle
Notes
The widest fan-out in the release: twelve fixed-corpus partitions are searched in parallel and every one of them must contribute a verified quotation before entity resolution can commit to a single candidate. Answer and provenance run as an auditable pair, and twelve revision edges let a failed audit push the workflow back into the partition layer.
Workflow topology
P01 P02 P03 P04 P05 P06 P07 P08 P09 P10 P11 P12 entity resolution answer agent provenance auditor result synthesizer workflow verifier
handoff flow revision (bounded re-entry)
04
Mathematical & Structured Reasoning

HepLean ElevenPlane proof-dependency workflow

v15_family_0682_989ef425b8d7 16 nodes · 40 edges math-hep-0082 Lean 4 · HepLean-v4.7
Global goal (excerpt)
Maintain the fixed Lean project HepLean-v4.7 at commit 7448822afced644bd61f0bdcf4dc438f455f9890 with toolchain leanprover/lean4:v4.7.0. Treat the 15 declarations below as one real proof-dependency workflow. Independent declarations may be assigned to separate reasoning agents. A downstream agent may receive only a successfully compiled ProofReceipt whose theorem ID, statement hash, repository commit and toolchain match its declared incoming edge. Do not change signatures and do not use sorry, ad
Agent roles
reasoning_agent ×15 workflow_verifier
Artifact types on handoff edges
proof_receipt verification_receipt
Notes
Ten independent isSolution_f* obligations fan in to sum_part, then grav, and a further pair feeds f_zero. Every handoff is a ProofReceipt gated on theorem ID, statement hash, repository commit and toolchain — version-aware memory in its strictest form.
Workflow topology
f0 f1 f2 f3 f4 f5 f6 f7 f8 f9 sum_part grav f10 only_if_zero f_zero Lean release
handoff flow revision (bounded re-entry)
05 — Construction

How a source workflow becomes a memory experiment

Three construction modes, one revision contract, and domain-native execution. Topology is a reporting and stress dimension — not an asserted cause of failure without matched interventions.

Construction pipeline
STEP 01
Source record
Toolathlon / SWE-bench / BrowseComp+ / HepLean
STEP 02
Induced or native workflow
400 induced · 200 native · 200 dependency subgraph
STEP 03
Typed handoff graph
Artifact types + version pins on every edge
STEP 04
Revision contract
semantic_bounded_v19 — bounded revision loops
STEP 05
Control + pollution pair
800 control · 800 matched polluted instances
Topology motifs across the 800 families
chain
800 / 800
split_join
800 / 800
reuse
799 / 800
overlap
787 / 800
bounded_revision_loop
418 / 800
Problems
800
complete
One complete set — no held-out test/dev split.
Construction modes
source_grounded_induced_workflow400
source_native_workflow200
source_native_dependency_subgraph200
Revision contract
Mode semantic_bounded_v19. Bounded revision loops let failed receipts send control back one step — matching how real workflows recover without unbounded replanning.
06 — Evaluation

Headline metrics

Two standard measures and two memory-specific measures. Infrastructure failures are reported separately and never counted as wrong answers.

SR

Task Success Rate

End-to-end workflow success over the full 50-family domain denominator. Reported as completed / registered.

VNCR

Verified Native Completion Rate

Macro-mean of native_passed over nodes. Captures whether each node completed its own declared obligation under domain-native checks.

VHS

Verified Handoff Success

Micro-ratio of verified eligible handoffs — downstream success conditioned on successful mandatory predecessors. Proxy for required-memory availability.

ICS

Isolation Challenge Score

Success on polluted instances conditioned on matched control success. Blank when the denominator is below 10 — never reported as zero.

Cost
Tokens / Task — report input + output (median, P90)
Infra
Infrastructure failures reported separately from task failures
Blanking rule
— = not registered / denom < 10, never an invented zero
07 — Results

Leaderboard

Three backbone snapshots.

DeepSeek-V4-Flash
Qwen3.8-27B
GPT-6-Luna
Metric:
System Org. category Stateful Tool Repository Code Offline Retrieval Formal Mathematics Macro
SR = task success rate over 50-family domain denom VNCR = verified native completion rate (macro) VHS = verified handoff success (micro) ICS = isolation challenge score (provisional; blank if denom<10)
08 — Access

Get the dataset

CoMemBench v1 is published on HuggingFace — 800 complete problems across four domains, each with one matched polluted twin. No held-out split, no gate: download and run.

Download on HuggingFace datasets/SenLab/Comembench · v1 · CC BY 4.0
01

What’s in the release

workflow_families.jsonl (800 problems), node_contexts.jsonl (8975 memory scopes), retrieval_corpora.jsonl & math_cases.jsonl, topology descriptors, the instance indexes, and the JSON schemas.

02

Load in three lines

Plain JSON Lines — or use the datasets loader. Each family is a layered DAG of worker nodes with typed handoffs and a bounded revision contract.

03

Score against held-out oracles

Evaluation oracles and pollution plans stay held out. SR / VNCR / VHS / ICS are computed under the published blanking and denominator rules — infra failures never count as wrong answers.

09 — Citation

Cite CoMemBench

If you use CoMemBench, please cite the following work. A PDF of the manuscript is at comembench.world/paper/.

@misc{zhao2026comembench,
  title        = {{CoMemBench}: Benchmarking Collaborative Memory
                  Boundaries across Multi-Agent Workflow Topologies},
  author       = {Zhao, Sen and Kong, Ruiqi and Zhang, Zuyu and
                  Shen, Lifeng and He, Xinyu and Zou, Ding and
                  Zhang, Xu and Zhang, Qinghua},
  year         = {2026},
  howpublished = {\url{https://comembench.world/paper/}},
  note         = {Manuscript under review}
}