Datasets:

caiovicentino1
/

agentlife-benchmark

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
query_id: string
question: string
qtype: string
day_asked: int64
seq: int64
hops: int64
fact_keys: list<item: string>
  child 0, item: string
answer_display: string
accepted: list<item: string>
  child 0, item: string
rejected: list<item: string>
  child 0, item: string
heat: string
n_updates_before: int64
config: struct<preset: string, n_days: int64, n_persons: int64, n_projects: int64, events_min: int64, events (... 272 chars omitted)
  child 0, preset: string
  child 1, n_days: int64
  child 2, n_persons: int64
  child 3, n_projects: int64
  child 4, events_min: int64
  child 5, events_max: int64
  child 6, hot_fraction: double
  child 7, hot_pick_day: int64
  child 8, multihop_min_gap: int64
  child 9, quotas: struct<current_stable: int64, current_updated: int64, current_volatile: int64, point_in_time: int64, (... 69 chars omitted)
      child 0, current_stable: int64
      child 1, current_updated: int64
      child 2, current_volatile: int64
      child 3, point_in_time: int64
      child 4, multihop_2: int64
      child 5, multihop_3: int64
      child 6, revoked: int64
      child 7, absent: int64
  child 10, seed: int64
hot_set: list<item: string>
  child 0, item: string
est_tokens: int64
warnings: list<item: string>
  child 0, item: string
counts: struct<episodes: int64, online_queries: int64, final_queries: int64, final_by_type: struct<current_s (... 168 chars omitted)
  child 0, episodes: int64
  child 1, online_queries: int64
  child 2, final_queries: int64
  child 3, final_by_type: struct<current_stable: int64, current_updated: int64, current_volatile: int64, point_in_time: int64, (... 69 chars omitted)
      child 0, current_stable: int64
      child 1, current_updated: int64
      child 2, current_volatile: int64
      child 3, point_in_time: int64
      child 4, multihop_2: int64
      child 5, multihop_3: int64
      child 6, revoked: int64
      child 7, absent: int64
  child 4, facts: int64
to
{'config': {'preset': Value('string'), 'n_days': Value('int64'), 'n_persons': Value('int64'), 'n_projects': Value('int64'), 'events_min': Value('int64'), 'events_max': Value('int64'), 'hot_fraction': Value('float64'), 'hot_pick_day': Value('int64'), 'multihop_min_gap': Value('int64'), 'quotas': {'current_stable': Value('int64'), 'current_updated': Value('int64'), 'current_volatile': Value('int64'), 'point_in_time': Value('int64'), 'multihop_2': Value('int64'), 'multihop_3': Value('int64'), 'revoked': Value('int64'), 'absent': Value('int64')}, 'seed': Value('int64')}, 'counts': {'episodes': Value('int64'), 'online_queries': Value('int64'), 'final_queries': Value('int64'), 'final_by_type': {'current_stable': Value('int64'), 'current_updated': Value('int64'), 'current_volatile': Value('int64'), 'point_in_time': Value('int64'), 'multihop_2': Value('int64'), 'multihop_3': Value('int64'), 'revoked': Value('int64'), 'absent': Value('int64')}, 'facts': Value('int64')}, 'est_tokens': Value('int64'), 'hot_set': List(Value('string')), 'warnings': List(Value('string'))}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 489, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              query_id: string
              question: string
              qtype: string
              day_asked: int64
              seq: int64
              hops: int64
              fact_keys: list<item: string>
                child 0, item: string
              answer_display: string
              accepted: list<item: string>
                child 0, item: string
              rejected: list<item: string>
                child 0, item: string
              heat: string
              n_updates_before: int64
              config: struct<preset: string, n_days: int64, n_persons: int64, n_projects: int64, events_min: int64, events (... 272 chars omitted)
                child 0, preset: string
                child 1, n_days: int64
                child 2, n_persons: int64
                child 3, n_projects: int64
                child 4, events_min: int64
                child 5, events_max: int64
                child 6, hot_fraction: double
                child 7, hot_pick_day: int64
                child 8, multihop_min_gap: int64
                child 9, quotas: struct<current_stable: int64, current_updated: int64, current_volatile: int64, point_in_time: int64, (... 69 chars omitted)
                    child 0, current_stable: int64
                    child 1, current_updated: int64
                    child 2, current_volatile: int64
                    child 3, point_in_time: int64
                    child 4, multihop_2: int64
                    child 5, multihop_3: int64
                    child 6, revoked: int64
                    child 7, absent: int64
                child 10, seed: int64
              hot_set: list<item: string>
                child 0, item: string
              est_tokens: int64
              warnings: list<item: string>
                child 0, item: string
              counts: struct<episodes: int64, online_queries: int64, final_queries: int64, final_by_type: struct<current_s (... 168 chars omitted)
                child 0, episodes: int64
                child 1, online_queries: int64
                child 2, final_queries: int64
                child 3, final_by_type: struct<current_stable: int64, current_updated: int64, current_volatile: int64, point_in_time: int64, (... 69 chars omitted)
                    child 0, current_stable: int64
                    child 1, current_updated: int64
                    child 2, current_volatile: int64
                    child 3, point_in_time: int64
                    child 4, multihop_2: int64
                    child 5, multihop_3: int64
                    child 6, revoked: int64
                    child 7, absent: int64
                child 4, facts: int64
              to
              {'config': {'preset': Value('string'), 'n_days': Value('int64'), 'n_persons': Value('int64'), 'n_projects': Value('int64'), 'events_min': Value('int64'), 'events_max': Value('int64'), 'hot_fraction': Value('float64'), 'hot_pick_day': Value('int64'), 'multihop_min_gap': Value('int64'), 'quotas': {'current_stable': Value('int64'), 'current_updated': Value('int64'), 'current_volatile': Value('int64'), 'point_in_time': Value('int64'), 'multihop_2': Value('int64'), 'multihop_3': Value('int64'), 'revoked': Value('int64'), 'absent': Value('int64')}, 'seed': Value('int64')}, 'counts': {'episodes': Value('int64'), 'online_queries': Value('int64'), 'final_queries': Value('int64'), 'final_by_type': {'current_stable': Value('int64'), 'current_updated': Value('int64'), 'current_volatile': Value('int64'), 'point_in_time': Value('int64'), 'multihop_2': Value('int64'), 'multihop_3': Value('int64'), 'revoked': Value('int64'), 'absent': Value('int64')}, 'facts': Value('int64')}, 'est_tokens': Value('int64'), 'hot_set': List(Value('string')), 'warnings': List(Value('string'))}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

🧠 AgentLife — a self-validating benchmark for long-lived agent memory

Paper: The Ledger Knows What the Weights Know: Entity-Level Parametric Memory with Provenance Guarantees for LLM Agents — doi:10.5281/zenodo.21375428 (concept DOI, resolves to latest; this mirror ships the exact datasets evaluated in v2.1: 10.5281/zenodo.21396695) · Code + generator: https://github.com/caiovicentino/cls-ledger

AgentLife generates synthetic "lives" of a personal-assistant agent — a stream of dated episodes where facts are stated, updated, revoked, and interleaved with noise — followed by a final exam. The answer key is correct by construction: every answer derives from a programmatic temporal world state, never hand-annotation.

🎯 Why another memory benchmark

Each design principle answers a documented failure of existing agent-memory benchmarks (e.g. the LoCoMo audit: 6.4% of the answer key wrong, an LLM judge accepting ~63% of deliberately wrong answers):

principle	enforcement
Answer key correct by construction	a privileged oracle must score exactly 100% — CI test
Freshness is falsifiable	a deliberately stale oracle must score 0% on updated-fact probes — CI test
Exact, auditable scoring	every query ships `accepted` + `rejected` surface forms; word-boundary matching; no LLM judge
Typed errors	`correct / stale / hallucination / wrong_value / ambiguous / abstain / miss`
Instrumented usage	online queries mark facts hot/cold → consolidation-policy experiments
Deterministic	same preset+seed → byte-identical dataset across processes and hash seeds — CI test

📊 Query types

current_stable · current_updated (old values rejected) · current_volatile · point_in_time (value as of day D) · multihop_2/3 · revoked · absent (hallucination/abstention probe) · online — plus, in v2 datasets (*v5): quantitative induction (count / weekday-habit / trend over fact histories) and behavioral dispositions (rules declared once, scored at the final exam with no reminder), and an adversarial paraphrase protocol (final_queries_paraphrased.jsonl) that measures benchmark–system co-adaptation.

📁 Layout

13 datasets: S-1..S-5 (30-day lives), M-1..M-3 (90-day), M-9v5..M-12v5 (90-day with induction+disposition cells), L-1 (365-day). Each contains:

manifest.json
public/            # all a system may see
  stream.jsonl     # dated episodes (facts + noise)
  episodes.jsonl
  final_queries.jsonl
  final_queries_paraphrased.jsonl
private/           # scoring only — answer key, world state, usage stats

⚠️ private/ is for scoring only. A system under evaluation must never read it; the split is the benchmark's contract, not a secrecy mechanism.

🔬 Reference-system signature (M-1)

system	final acc	what it validates
oracle	100%	answer key + scorer, end to end
stale oracle	64.6%	0% on updated/volatile/revoked — rejected-sets catch staleness
null	9.8%	abstention floor (100% on `absent` only)

🚀 Use

git clone https://github.com/caiovicentino/cls-ledger
# regenerate any dataset bit-for-bit (deterministic per seed):
PYTHONPATH=. python3 -m agentlife.generate --preset M --seed 1 --out data/M-1
# validate the mirror: the oracle must print 100%
PYTHONPATH=. python3 -m agentlife.harness.run_eval --data data/M-1 --system oracle

📖 Citation

@misc{vicentino2026ledger,
  author    = {Vicentino, Caio},
  title     = {The Ledger Knows What the Weights Know: Entity-Level
               Parametric Memory with Provenance Guarantees for LLM Agents},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21375428},
  url       = {https://doi.org/10.5281/zenodo.21375428}
}

🔗 Links

📄 Paper (v2.1): https://doi.org/10.5281/zenodo.21396695
💻 Code, generator, harness: https://github.com/caiovicentino/cls-ledger
📦 obsidian-ledger (the architecture on your own vault, any coding agent): https://github.com/caiovicentino/cls-ledger/releases/tag/v2.1
🧭 OpenInterpretability collection: https://huggingface.co/collections/caiovicentino1/openinterpretability

🙏 Acknowledgements

Experiments, code, and drafting carried out with the assistance of Claude (Anthropic). Every number in the paper maps to a JSON artifact in the repository (paper/REPRODUCING.md).

Downloads last month: 27

Total file size:

2.97 MB

Collection including caiovicentino1/agentlife-benchmark

OpenInterpretability

Collection

Agent-safety interpretability arc, first public Qwen3.6 SAEs, the probe/guard suite, and the agent-trajectory benchmarks behind them. openinterp.org • 37 items • Updated 13 days ago