README.md

---
pretty_name: ISETrace Memory Query Corpus
license: cc-by-4.0
language:
- en
size_categories:
- 1K<n<10K
task_categories:
- question-answering
tags:
- agent-memory
- information-retrieval
- tool-use
- execution-provenance
configs:
- config_name: queries
  data_files:
  - split: full
    path: query-authoring/isetrace-v7-raw.jsonl
- config_name: metadata
  data_files:
  - split: full
    path: query-authoring/isetrace-v7-raw.jsonl.metadata.jsonl
---

# ISETrace Memory Query Corpus

This repository contains the frozen natural-language memory-query corpus used for exact-span retrieval experiments over [ISETrace](https://huggingface.co/datasets/valiere/ISETrace), an execution-grounded corpus of operating-system agent trajectories.

The release contains 6,597 English queries. Each query asks for information recoverable from one trajectory and identifies minimal answer-bearing quotes in stable source sections. Query records and authoring metadata are kept separate so metadata such as query type is not accidentally treated as a retrieval feature.

## Files

| File | Records | Description |
|---|---:|---|
| `query-authoring/isetrace-v7-raw.jsonl` | 6,597 | Frozen four-field memory-query records |
| `query-authoring/isetrace-v7-raw.jsonl.metadata.jsonl` | 6,597 | Query-to-trajectory and query-type metadata |

The query file has SHA-256 `3add2685ea45525d84ba9ed797d658df2d23c29ae0368699a69c348dea39d7f8`.
The metadata file has SHA-256 `30db73bda691d68c6e62495c52d08854ef2007a8a035d0f3e85e128b874c28a0`.

## Query schema

Each line in `isetrace-v7-raw.jsonl` has exactly four fields:

```json
{
  "id": "query:...",
  "text": "[I1 | user_intent]...\n\n[A1 | tool_call | ...]...\n\n[E1 | tool_output | ...]...",
  "query": "What information should be recovered from this execution?",
  "gold": [
    {
      "source": "E1",
      "quote": "an exact answer-bearing substring"
    }
  ]
}
```

`I*` identifies a user-intent section, `A*` a tool-call argument section, and `E*` a tool-output section. Every gold quote occurs exactly once in its named source section. Downstream preparation resolves these quotes to exact character spans in the corresponding ISETrace trajectory.

## Metadata schema

Each aligned metadata line contains:

```json
{
  "query_id": "query:...",
  "task_key": "T-...",
  "trajectory_id": "traj_...",
  "memory_mode": "linked_recall"
}
```

The frozen corpus contains 2,134 `direct_recall`, 3,004 `linked_recall`, and 1,459 `multi_fact_recall` records. `memory_mode` supports audit and grouped reporting; it is not a retrieval feature or relevance label.

## Source data and split construction

The source trajectories are the `trajectories` configuration of:

- Repository: `valiere/ISETrace`
- Revision: `e40e04d41c04e4eb4bae181ebdd41b61c688081b`
- License: CC BY 4.0

The query corpus is intentionally released as one `full` split. The experiment code assigns whole trajectories, never individual queries, using split seed 13. The frozen full-profile allocation contains:

| Experiment split | Trajectories | Queries |
|---|---:|---:|
| train | 2,410 | 3,894 |
| validation | 352 | 580 |
| test | 1,207 | 2,000 |

This trajectory-level assignment prevents a trajectory from appearing in more than one experiment split. The remaining authored records belong to trajectories outside the fixed allocation.

## Loading

```python
from datasets import load_dataset

queries = load_dataset(
    "HazeLocus/ISETrace-Memory-Queries",
    "queries",
    split="full",
)
metadata = load_dataset(
    "HazeLocus/ISETrace-Memory-Queries",
    "metadata",
    split="full",
)
trajectories = load_dataset(
    "valiere/ISETrace",
    "trajectories",
    split="train",
    revision="e40e04d41c04e4eb4bae181ebdd41b61c688081b",
)
```

The accompanying experiment code downloads both repositories automatically at their pinned revisions when the expected local files are absent.

## Construction and validation

Queries were authored by a language model from bounded source packets selected from complete trajectories. Local validation enforces structured output, valid and unique source handles, exact quote grounding, unique query IDs, and exact alignment between query and metadata IDs.

A stratified sample of 500 queries underwent dual-model review, adjudication, and complete human validation. The full 6,597-query corpus was not independently human re-annotated. Users should account for possible query-type noise, lexical cues from authoring, and omitted alternative evidence. This release is a frozen research artifact rather than a claim that every query is human-verified.

## Intended use

The corpus supports research on retrieval from completed tool-use trajectories, particularly exact-span evidence recovery under fixed context budgets. It does not evaluate online memory updates, cross-session memory, or agent-to-agent communication. ISETrace personas are synthetic; execution paths and command outputs are preserved from the upstream execution-grounded dataset.

## License and attribution

This derived corpus includes excerpts from ISETrace and is released under CC BY 4.0. Users must attribute the upstream ISETrace dataset and indicate that memory queries and exact-quote annotations are derived material.

```bibtex
@misc{isetrace2026,
  title        = {From Intent to Trajectory: Execution-Grounded Multi-Turn Data Synthesis for OS Agents},
  author       = {Valiere01},
  year         = {2026},
  howpublished = {\url{https://github.com/Valiere01/ISE-Trace}}
}
```