pretty_name: AtMem Memory Integrity and Continuity Evidence
license: apache-2.0
language:
- en
tags:
- agent-memory
- llm-agents
- benchmark
- evaluation
- provenance
- crash-recovery
- tabular
configs:
- config_name: integrity_categories
data_files:
- split: test
path: data/integrity_categories.csv
- config_name: integrity_metrics
data_files:
- split: test
path: data/integrity_metrics.csv
- config_name: continuity_no_fault
data_files:
- split: test
path: data/continuity_no_fault.csv
- config_name: continuity_fault
data_files:
- split: test
path: data/continuity_fault.csv
- config_name: continuity_acceptance
data_files:
- split: test
path: data/continuity_acceptance.csv
- config_name: continuity_live_pilot
data_files:
- split: test
path: data/continuity_live_pilot.csv
AtMem memory integrity and continuity evidence
This dataset accompanies Beyond Recall Accuracy: Evaluating Integrity and Crash Continuity in Persistent Memory for Tool-Using Language Agents by Javad Taghia, AtMem.Ai Lab.
Open the interactive results Space · Read the PDF · AtMem source · External benchmark submission
What is measured
The publication separates two questions:
- Memory integrity: can untrusted observations, summaries, repetition, or prior agent outputs silently acquire factual or procedural authority?
- Operational continuity: after a crash, can an agent distinguish a saved success, a queryable external effect, a valid idempotent retry, and an outcome that must stop for confirmation?
Headline results
| Evidence | Result | Correct interpretation |
|---|---|---|
| Integrity trials | 700 | 400 PASS, 300 NOT_REPRESENTABLE, 0 FAIL, 0 ERROR |
| Factual contamination | 0/300 | Within represented scored probes |
| Secret retention | 0/100 | Within canonical benchmark scoring surfaces |
| Taint preservation | 100/100 | Direct-parent derivations in Category D |
| No-fault parity | 18/18 requests in every arm | One recorded-response trajectory |
| Repeat request after interrupted commit | baseline 1, AtMem 0, AtFlows 1, both 0 | AtMem avoided an uncertain redispatch in one fault cell |
NOT_REPRESENTABLE is not a pass or failure. The harness supplies caller labels,
not AtMem-issued review authority, for Categories A, H, and I. Category D uses
native-record-ID matching and remains pending upstream interpretation.
Repository layout
data/: small CSV tables rendered by the Dataset Viewer.evidence/integrity/: exact 700-trial integrity result bundle.evidence/continuity/: exact installed acceptance, replay, fault, and live-pilot bundles.paper/: PDF, LaTeX source, and bibliography.provenance/: top-level checksums and publication manifest.
Limitations
This is not a general security, compliance, retrieval-quality, or exactly-once
execution claim. The fresh four-arm pilot contains one exposed task per arm and
cannot support causal inference. The external integrity pull request is open and
is not an accepted upstream leaderboard entry. See the paper and bundled
LIMITATIONS.md for the complete claim boundary.
Paper
The bundled PDF is the canonical paper for this Hugging Face release. No arXiv identifier has been assigned. If the paper is submitted to arXiv later, this card can link the resulting Hugging Face Paper page without changing the dataset.
Citation
@article{taghia2026beyond,
title={Beyond Recall Accuracy: Evaluating Integrity and Crash Continuity in Persistent Memory for Tool-Using Language Agents},
author={Taghia, Javad},
year={2026},
note={AtMem.Ai Lab; Hugging Face research release}
}
OpenClaw lost-acknowledgement supplement (27 September 2026)
Read the paper · LaTeX source · Evidence and reproduction
When the Write Lands but the Acknowledgement Is Lost: A Reproducible Memory-Retry Test in OpenClaw, Javad Taghia, AtMem.Ai Lab.
Five isolated bridge crash trials, three real OpenClaw gateway crash trials, and one no-fault gateway control passed. After the completed write's response was withheld and the backend killed, the retry returned the original record as already_stored; read-back showed the complete original fact and one active record. Gateway model responses were deterministic and local. This tests exact-content retry after a completed write, not partial transactions or power loss. The supplement preserves two earlier CLI setup failures separately and does not change the earlier study's results.