MemArena-L v1.0.0 — Dataset Card

MemArena: An Ego-Centric Benchmark for Multi-User Agentic Memory on On-Device Open-Weight Models. Anonymous Authors, 2026. NeurIPS 2026 Evaluations & Datasets Track (double-blind review).

1. Summary

MemArena-L is the Large split of the MemArena benchmark for evaluating long-horizon, multi-user agentic memory. A single coherent multi-agent world of 50 agents over 15 simulated days is generated by the MASim simulator, with ego-centric projections restricting each user to the sessions they actually witnessed. Evaluation probes span six paper-level dimensions implemented as nine evaluated internal task files in three categories: Recall (episodic), Reasoning (factual QA, temporal, cross-session), and Trustworthiness (abstention, permission-aware access).

Aspect	Value
Agents	50
Simulated days	15
Sessions	13,343
Conversational turns	137,279
Corpus tokens	10.3 M
Ego-centric avg sessions per user	401
Ego-centric avg turns per user	4,186
Probe families	9 evaluated internal files (d1--d8, d10)
Total eval instances	1,579
Compressed size	42 MB
License	CC BY 4.0
Version	1.0.0 (released 2026-04-24)

Corpus tokens are counted over each dialogue turn's text field with the Qwen/Qwen3-8B tokenizer; 10.3 M rounds the exact count of 10,305,361.

2. Intended uses

• Benchmark memory-augmented LLM agents on long-horizon ego-centric recall
• Compare backends (vanilla, RAG, oracle, structured memory like Memobase / MemOS)
• Study privacy/permission behavior under multi-user context
• Calibrate on-device open-weight models (Qwen3-0.6B/8B/32B, Llama-3.2-3B, Mistral-7B)

Not intended for: training chat models on MemArena text (text is synthetic and stylistically narrow); profiling real users (there are no real users); benchmarking retrieval speed on large-scale web corpora.

3. Files

The Hugging Face release exposes these files at the repository root. Tarball releases contain the same file tree in memarena-l-v1.0.0.tar.gz.

Path	Format	Description
agents_personas.jsonl.gz	gzip JSONL	50 rows: {agent_id, persona, social_neighbors, person_impressions}
agent_schedules.jsonl.gz	gzip JSONL	Per-agent daily activity schedules across 15 days
corpus_sessions.jsonl.gz	gzip JSONL	13,343 rows: session with {session_id, participants, turns, start_time, end_time, triggering_events, location_id, group_id, role_name, modality, interest_domain, session_type, metadata} — each turn has {turn_id, speaker_id, listener_id, text, timestamp, triggering_event, entities_mentioned, extracted_facts}
events.jsonl	JSONL	75 world events: {event_id, timestamp, event_type, content, visibility_mask, category, location_id, interest_domain, metadata}
groups.json	JSON	21 social groups with member lists
locations.json	JSON	67 locations with coordinates / descriptions
graph_edges.jsonl	JSONL	Social graph {source, target, weight} (1,225 edges)
ego_projections.json	JSON	Per-agent ego-centric session summaries (avg 401 sessions / 4186 turns)
ego_session_map.json	JSON	Mapping {agent_id → [session_ids witnessed]}
eval_instances/{d1,d2,d3,d4,d5,d6,d7,d8,d10}_*.jsonl	JSONL	9 evaluated internal task files, 1,579 instances total
day_checkpoint.json	JSON	Per-day generation checkpoint (for reproducibility)
pipeline_report.json	JSON	Authoritative generation statistics and config

3.1 Probe family schema (eval_instances/d*.jsonl)

Common fields:

instance_id          string
dimension            enum {d1_conflict, d2_anaphora, d3_confabulation, d4_permission,
                           d5_cloze, d6_metadata, d7_qa, d8_temporal,
                           d10_counterfactual}
query                string                — the user-facing prompt
ground_truth         object                — dimension-specific schema (see below)
difficulty           enum {easy, medium, hard}
asker_agent_id       string                — who asks
answerer_agent_id    string                — who is expected to answer
ego_agent_id         string                — whose ego-centric view this probes
evidence_session_ids array<string>         — sessions providing evidence
context_tier         enum {single_session, cross_session, full_ego}
metadata             object                — free-form, includes query_timestamp, query_modality, etc.

Dimension-specific ground_truth keys (brief):

Dim	Key extras
d1_conflict	source_agent, conflicting_agent, original_session, conflicting_session, original_fact, contradictory_fact, original_detail, changed_detail
d2_anaphora	antecedent_span, anaphor_span, target_entity
d3_confabulation	plausible_but_absent_fact, near_miss_evidence
d4_permission	permission_granted, permission_revoked, allowed_listeners, restricted_listeners
d5_cloze	target_span, distractors, answer
d6_metadata	target_metadata_field, answer
d7_qa	answer, source_session, requires_temporal
d8_temporal	temporal_anchor, temporal_offset, answer
d10_counterfactual	counterfactual_branch, base_fact, answer

3.2 Counts per probe family

Dim	Count
d1_conflict	170
d2_anaphora	170
d3_confabulation	170
d4_permission	200
d5_cloze	198
d6_metadata	169
d7_qa	170
d8_temporal	162
d10_counterfactual	170
Total	1,579

4. Data generation (provenance)

MemArena-L was generated by MASim (this repository), a multi-agent simulation pipeline. High-level stages:

1. Social graph construction — Watts-Strogatz with 3 Dunbar layers (5, 15, 50) on 50 agents; 1,225 edges; fully connected; seed 42.
2. Persona synthesis — personas drawn from configs/reviewed_personas_l.jsonl (reviewed for demographic/occupation balance).
3. Schedule generation — 18 sessions/agent/day target; mixture of facet-routine (40%), event-triggered (15%), encounter (25%), remote (20%).
4. Session dialog synthesis — Qwen3 (served via vLLM) generates turns conditioned on participants' personas, location, role, recent memory, and triggering events. Temperature 0.7, max 512 tokens per turn, min 4 / max 12 turns per session.
5. Injection passes — 20 deliberate conflicts (d1), 294 permission-marked utterances (d4), 196 autonomous-privacy cases; counterfactual, temporal, paraphrase, and perspective hardeners for probe templates.
6. Ego projection — for each agent u_i, retain only sessions where u_i ∈ participants.
7. Probe extraction — nine evaluated internal task extractors over the full corpus, each producing 162--200 instances.
8. Query hardening — paraphrase, temporal, perspective, and counterfactual hardening, bounded at 30% max lexical overlap with source text.

Generator model: Qwen3 served locally. Judge (at eval time, not baked in): GPT-4o-mini.

5. Responsible AI disclosures

5.1 Known limitations (rai:dataLimitations)

• Synthetic text only — no real human conversations; distributional coverage of real-world dialog is limited.
• English only — all turns are in English; no code-switching or non-English evaluation.
• Single simulator family — the base LM for generation is Qwen3; stylistic artifacts of that model family are present in the text.
• Small population — 50 agents; not suitable for population-level statistical claims.
• 15-day horizon — does not probe year-long memory.
• Bounded probe templates — each dimension uses ≤ N template families; probe diversity is simulator-bounded.
• No multimodal content in v1.0.0 — enable_image and enable_audio were off; image/audio hooks exist in the schema but carry no payload.

5.2 Known biases (rai:dataBiases)

• Generator-model bias — phrasing, vocabulary, and discourse style reflect Qwen3 priors.
• Persona-sampling bias — personas were filtered for diversity but drawn from a reviewer pool that over-represents STEM occupations and urban residency.
• English-centric cultural priors — events, holidays, and idioms skew Western-urban.
• Topic imbalance — interest_domain distribution is not uniform; some domains (tech, arts) are over-represented relative to others (agriculture, trades).
• Graph topology — Watts-Strogatz with k=4–6; small-world, but not heavy-tailed like real social networks.

5.3 Personal & sensitive information (rai:personalSensitiveInformation)

• No real personal information. All agent names, locations, and personas are synthetic.
• No protected-category PII (medical records, government IDs, financial accounts) is intentionally simulated.
• The benchmark does simulate permission-marked utterances (d4) and autonomous-privacy disclosures (e.g., mentions of made-up health events, made-up political views) for the purpose of evaluating permission-aware withholding. These are fictional and do not correspond to any real person.
• Users of this dataset must not treat simulated personas as representative of any real demographic group.

5.4 Use cases (rai:dataUseCases)

• Benchmark memory-augmented agents
• Ablate retrieval/summarization/structured-memory backends
• Study ego-centric projection vs. omniscient-context baselines
• Test abstention / permission-aware withholding
• Compare on-device (open-weight ≤ 32B) models at matched compute

5.5 Social impact (rai:dataSocialImpact)

Positive: establishes an ecologically calibrated, privacy-aware evaluation target for on-device personal agents, which shifts incentives toward open-weight models and away from centralized cloud-only deployments.

Negative / misuse risk: the d4_permission probes encode realistic permission-violation patterns. Researchers fine-tuning on these splits as training data (rather than evaluation) could inadvertently teach models to reproduce permission violations. Use for evaluation only.

5.6 Synthetic data flag

rai:hasSyntheticData = true — the entire corpus is simulator-generated.

5.7 Derived from (prov:wasDerivedFrom)

• Persona pool: MASim/configs/reviewed_personas_l.jsonl (internal reviewed pool, first release with v1.0.0)
• Activity topics: MASim/data/activity_topics.yaml
• Location ontology: MASim/data/location_cache/

5.8 Generated by (prov:wasGeneratedBy)

MASim pipeline, commit pinned in pipeline_report.json → metadata.commit_sha. Run configuration archived in pipeline_report.json → config (full dump).

6. Baseline results

Aggregate baseline scores for 5 on-device models × 5 backends × 4 seeds are reported in the companion MemArena paper. Per-instance model outputs and judge rationales are not distributed; users can reproduce them with the evaluation code and the cleaned corpus shipped here.

7. Citation

@inproceedings{memarena2026,
  title     = {MemArena: An Ego-Centric Benchmark for Multi-User Agentic Memory on On-Device Open-Weight Models},
  author    = {Anonymous Authors},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  note      = {Datasets and Benchmarks Track},
}

8. License

CC BY 4.0 — see LICENSE. You may share and adapt with attribution.

9. Changelog

v1.0.0 (2026-04-24)

• Initial release: 50 agents × 15 days; 1,579 evaluated instances across nine internal task files.
• Aggregate baseline scores reported in the paper; per-instance outputs are not distributed.
• On top of the raw Qwen3 simulation output, the authors ran an editorial pass that rewrote 1,198 fields (775 corpus sessions, 15 agent personas, and 3 probe items) to enforce English-only dialogue. The pass preserves meaning, tone, emoji, and formatting; it modifies ~0.02% of corpus tokens. Intentional non-ASCII patterns retained by the authors: Greek quoted phrases paired with in-character English glosses; one Ukrainian proverb (quoted, transliterated, and translated in the same utterance); and a small number of multilingual kaomoji.