Yummytanmo/SubtleMemory

Dataset

SubtleMemory

0

2 commits

2 linked in READMEs

updated Jun 9, 2026

See the code

README

SubtleMemory

SubtleMemory is a synthetic benchmark for fine-grained relational memory discrimination in long-horizon AI agents. It tests whether memory systems can preserve and use the relation among target-relevant memories rather than only retrieve isolated facts.

This dataset release preserves the original repository file structure:

data/subtlememory/
  persona_0/
    bench_instances.json
    history_sessions.json
  ...
  persona_9/
    bench_instances.json
    history_sessions.json

The Hugging Face Dataset Viewer is configured with persona-level splits:

  • bench_instances: one split per persona, containing benchmark cases and QA instances.
  • history_sessions: one split per persona, containing the chronological conversation sessions.

Counts

SplitBench instancesHistory sessions
persona_0102225
persona_1122266
persona_2122266
persona_3118257
persona_4122261
persona_5106226
persona_690199
persona_792193
persona_8100214
persona_9116257
total10902364

The benchmark contains 1,522 QA instances across the 1,090 benchmark cases.

Fields

bench_instances.json contains:

  • case_id, instance_id, persona_id, source, and topic
  • relation_type and relation_subtype
  • case, facts, and session_ids
  • qas, where each QA item contains query, correct_answers, and incorrect_answers
  • persona_str, the synthetic persona profile serialized as JSON text

history_sessions.json contains:

  • session_id, case_id, persona_id, source, order, and timestamp
  • conversation_type, conversation_flow, and persona_signal_level
  • history, a list of {role, content} conversation turns

Synthetic Data Notice

The data is synthetic benchmark data generated for research use. Persona identities, histories, contact details, account numbers, file paths, and other sensitive-looking values are fabricated and should not be interpreted as real personal data.

Usage

from datasets import load_dataset

bench_p0 = load_dataset(
    "Yummytanmo/SubtleMemory",
    "bench_instances",
    split="persona_0",
)

history_p0 = load_dataset(
    "Yummytanmo/SubtleMemory",
    "history_sessions",
    split="persona_0",
)

License

Apache-2.0.

benchmark
long-horizon-agents
memory
relation-discrimination
synthetic

Yummytanmo/SubtleMemory

Dataset

SubtleMemory

0

2 commits

2 linked in READMEs

updated Jun 9, 2026

See the code

README

SubtleMemory

SubtleMemory is a synthetic benchmark for fine-grained relational memory discrimination in long-horizon AI agents. It tests whether memory systems can preserve and use the relation among target-relevant memories rather than only retrieve isolated facts.

This dataset release preserves the original repository file structure:

data/subtlememory/
  persona_0/
    bench_instances.json
    history_sessions.json
  ...
  persona_9/
    bench_instances.json
    history_sessions.json

The Hugging Face Dataset Viewer is configured with persona-level splits:

  • bench_instances: one split per persona, containing benchmark cases and QA instances.
  • history_sessions: one split per persona, containing the chronological conversation sessions.

Counts

SplitBench instancesHistory sessions
persona_0102225
persona_1122266
persona_2122266
persona_3118257
persona_4122261
persona_5106226
persona_690199
persona_792193
persona_8100214
persona_9116257
total10902364

The benchmark contains 1,522 QA instances across the 1,090 benchmark cases.

Fields

bench_instances.json contains:

  • case_id, instance_id, persona_id, source, and topic
  • relation_type and relation_subtype
  • case, facts, and session_ids
  • qas, where each QA item contains query, correct_answers, and incorrect_answers
  • persona_str, the synthetic persona profile serialized as JSON text

history_sessions.json contains:

  • session_id, case_id, persona_id, source, order, and timestamp
  • conversation_type, conversation_flow, and persona_signal_level
  • history, a list of {role, content} conversation turns

Synthetic Data Notice

The data is synthetic benchmark data generated for research use. Persona identities, histories, contact details, account numbers, file paths, and other sensitive-looking values are fabricated and should not be interpreted as real personal data.

Usage

from datasets import load_dataset

bench_p0 = load_dataset(
    "Yummytanmo/SubtleMemory",
    "bench_instances",
    split="persona_0",
)

history_p0 = load_dataset(
    "Yummytanmo/SubtleMemory",
    "history_sessions",
    split="persona_0",
)

License

Apache-2.0.

benchmark
long-horizon-agents
memory
relation-discrimination
synthetic