Previously — Benchmark Data

Seeded conversation datasets for Previously, a personal AI commander platform where cloud agents work while you're away and results are waiting when you return.

Each persona in this repository is a complete, ready-to-use episodic memory store — 30 multi-turn conversations spanning 3–4 years, with rich metadata designed for the Previously memory architecture.

Quick Start

With Previously deployed on Vercel, set one environment variable:

DEMO_MODE=true

The app reads persona data directly from this repository via raw.githubusercontent.com — no GitHub token, no API key, no authentication required. Choose a persona from the UI and start chatting immediately.

Source & Attribution

All personas are derived from WorldMemArena, a large-scale multimodal memory benchmark introduced in:

Liu et al., "WorldMemArena: Evaluating Multimodal Agent Memory Through Action–World Interaction", arXiv:2605.29341, 2026.

WorldMemArena contains 461 samples across 8,489 sessions with 59,239 conversation turns and 40,194 annotated memory points, spanning lifelong personal evolution and agentic task execution. The 20 personal samples used here average 30 sessions each (~57K words per persona, comparable to a short novel).

We are grateful to the WorldMemArena authors (Chengzhi Liu, Yuzhe Yang, Xin Eric Wang, and the UCSB AI team) for creating and releasing this benchmark under CC BY-NC 4.0, which made this seeded dataset possible.

What's Inside

20 Complete Personas

Each persona is a fictional individual with a coherent multi-year life story — career progression, family dynamics, health challenges, financial decisions, hobbies, and personal growth — told through 30 natural conversation sessions with an AI assistant.

Persona	Person
personal_01	Marisa Elaine Cardenas — field supervisor, county restoration
personal_02	Lucia Elena Montoya — court interpreter, civic reform advocate
personal_03	Derek Anthony Kubik — journeyman wireman, IBEW Local 5
personal_04	Tasha Nicole Whitfield — bereavement coordinator, Raleigh
personal_05	Nathaniel James Hollis — transit scheduling analyst, COTA
personal_06	Hannah Louise Mercer — registrar, Providence Museum of Art
personal_07	Evan Michael Dorsey — procurement analyst, school district
personal_08	Jermaine Allen Booker — safety trainer, Feeders Supply
personal_09	Nadia Samira Rahal — intake caseworker, family services
personal_10	Kelli Ann Brotherton — revenue-cycle lead, Emmett ID
personal_11	Monica Renee Halverson — housing code supervisor, Milwaukee
personal_12	Ariana Celeste Batiste — archivist, oral history program
personal_13	Danielle Marie Rosas — nutrition coordinator, school district
personal_14	Caleb Martin Hebert — floodplain outreach specialist, Harris County
personal_15	Park Jae-hyun — dispatch coordinator, Busan
personal_16	Saira Nadeem Hussain — operations lead, Westgate Pharmacy, Leeds
personal_17	Colleen Margaret Sweeney — education manager, Fireman's Hall Museum
personal_18	Jordan Elise Petrov — operations planner, Alder Street Roasting
personal_19	Nakamura Ayaka — operations coordinator, Hakata ferry terminal
personal_20	Eric Matthew Lang — cold-chain coordinator, Freestore Foodbank

Previously Memory Format

Each persona directory is a self-contained episodic memory store:

personal_14/                       # Caleb Martin Hebert
├── episodic/
│   ├── slices/
│   │   ├── 2025/                  # Year
│   │   │   ├── 01/                # Month
│   │   │   │   ├── 08/            # Day
│   │   │   │   │   └── 1130.md    # Time slice (YAML frontmatter + turns)
│   │   │   │   └── ...
│   │   │   └── _index.json        # Monthly index of all slices
│   │   └── ...
│   └── strands.json               # Keyword → slice-path index
├── user/
│   └── profile.md                 # Persona profile
└── quality-report.json            # Metadata quality audit

Key concepts:

• Time Slice — A single conversation session. YAML frontmatter carries slice_id, focus, summary, tags, emotional_tone, open_loops, and decisions. The body records each turn (## Turn N — timestamp (role)) as structured Markdown.

• Monthly Index (_index.json) — Lightweight scan target for the Flash recall system. Lists every slice in a month with its metadata, so the system can find relevant conversations without reading full files.

• Strands (strands.json) — A keyword → slice-path index. Each tag on a slice weaves a strand through time: "flood-mitigation-outreach" → ["2025/01/08/1130", "2025/04/09/1010", ...]. This is the thin, lossless semantic layer over the chronological slices.

• Quality Report — Generated by Claude Haiku during enrichment. Flags metadata issues found and fixed in the original WorldMemArena conversion.

Data Quality

Every slice in this repository has been reviewed by Claude Haiku 4.5:

• Tags: Replaced generic regex-generated categories (work, family, health) with 4–8 specific, searchable kebab-case tags per slice (flood-mitigation-outreach, mother-health-appointment, post-approval-retention-routine)
• Emotional tone: Corrected from universal "neutral" to accurate per-slice values
• Focus & Summary: Fixed truncated or quote-fragment metadata
• Open loops: Added unresolved action items missing from the original conversion

This produces ~200–260 unique strands per persona, enabling precise conversational recall.

How Previously Uses This Data

Previously uses a Flash/Pro split memory architecture:

1. Flash (fast, cheap model) scans monthly _index.json files and strands.json to identify relevant past conversations
2. Pro (main agent model) deep-reads specific time slices when Flash finds matches
3. Every chat turn is recorded as a new time slice, building the user's own memory on top of the seed data (when GitHub write access is configured)

This repository provides the seed data so users can experience the full memory system immediately, without waiting months for their own conversation history to accumulate.

Conversion

The raw WorldMemArena JSON files are converted to Previously format using scripts/batch-convert.mjs in the Previously repository. Metadata enrichment is applied via scripts/apply-enrichment.mjs using Claude Haiku.

To regenerate from source:

# 1. Download WorldMemArena personal samples from Hugging Face
pip install huggingface_hub
huggingface-cli download LCZZZZ/WorldMemArena --repo-type dataset \
  --local-dir ./WorldMemArena \
  --include "WorldMemArena/lifelong/personal/personal_*.json"

# 2. Copy JSON files to benchmark-data/_raw/

# 3. Convert
node scripts/batch-convert.mjs --raw ../benchmark-data/_raw --out ../benchmark-data

# 4. Enrich with Haiku (requires ANTHROPIC_API_KEY or use Claude Code agents)
# See scripts/apply-enrichment.mjs for details

A Note on Dates

The original WorldMemArena dataset places all conversation timelines in 2025–2028. Since Previously should never have "memories from the future," the dates in this repository have been shifted back 3 years — all slices now fall within 2022–2025.

To restore the original dates:

node scripts/shift-all-dates.mjs --reverse

License

CC BY-NC 4.0 — inherited from WorldMemArena.

This means you are free to share and adapt this data for non-commercial purposes with appropriate attribution. For commercial use, please refer to the original WorldMemArena dataset and license terms.

Related Projects

• Previously — The app that consumes this data
• WorldMemArena — The original benchmark dataset
• WorldMemArena Paper — arXiv:2605.29341