Jacqueline "Jackie" Oliver

From Guardian memory database · 34 days of continuous life capture · Last updated April 2026

Jacqueline "Jackie" Oliver is an American technology entrepreneur and the cofounder of Haptica, a company building haptic gloves and cameras for capturing physical robotics training data. She is based in the San Francisco Bay Area and is the sole user and developer of Guardian, a continuous life-capture system that records across four modalities: desktop screen capture, wearable camera video, ambient audio recording, and mobile phone activity.[1]

Oliver is simultaneously Guardian's creator, its only test subject, and its product designer—a triple role that makes the 34-day dataset captured between March and April 2026 both a technology prototype and an unusually granular document of one person's life.[2]

Identity

Jackie's full first name is Jacqueline. She uses "Jackie" universally in conversation and on social media. Her X (Twitter) handle is @lambdacalcc.[3]

She does not speak French and never sings; all lyrics that appear in audio transcripts are background music, not her voice.[4]

Career & Haptica

Haptica builds haptic gloves and wearable cameras to capture physical robotics training data. The thesis is that real-world physical demonstration data—humans performing tasks while wearing instrumented gloves—is the critical bottleneck for training robot foundation models, and that simulation alone cannot bridge the sim-to-real gap.[6]

Oliver describes her product philosophy as: "Make something incredible for a few people, not mediocre for everyone."[7] This principle appears to guide both Haptica's hardware strategy and Guardian's development.

As of April 2026, Haptica is in an early fundraising stage. Oliver has discussed seeking angel investors, particularly "tech nerds giving angel checks"[8], and has been actively researching AI and VC events in the Bay Area.[9]

The company maintains a private GitHub organization for its development work.[10]

Paul Han — Cofounder

Paul Han appears throughout the capture data as Jackie's closest collaborator. He is referenced in audio transcripts, Discord direct messages, browser activity (scheduling meetings together), and visual captures across dozens of days.[11] See Paul Han for full article.

Daily Life

Nocturnal Rhythm

Oliver operates on a nocturnal schedule. Her typical "day" begins in the late afternoon and extends deep into the night.[12] This means her activity straddles calendar boundaries—the evening of calendar day N is the start of her waking day N+1.

A characteristic night involves long, unbroken sessions at the Linux desktop: development sprints with Claude Code and Codex interspersed with research browsing, music listening, Discord conversations, and AI-assisted inquiry. She rarely spends more than 10–15 minutes on a single task before branching, a pattern her own terminal notes describe as "ambient intellectual grazing" and "build + absorb + drift + return."[13]

Night Drives

A recurring motif in the bodycam data is Jackie driving through San Francisco at night, listening to music. The GoPro captures city intersections at night, music playing through the speakers.[14] The drives sometimes precede or follow social events—meeting Paul, running errands, or just moving through the city.[15]

Domestic Environment

The bodycam captures domestic sounds: a clock ticking, a cat meowing.[16] The primary work environment is a Linux desktop workstation, with a secondary MacBook used for browsing and communication when mobile.

Relationships

The Founder's Mindset

The capture data reveals someone building with urgency and conviction. "Luckily it's a good product. Its value is quite self-evident. It's not a tough sale."[17]

Close Collaborators

See individual articles: Paul Han, Keoni, Gwern. Other recurring contacts in the capture data include members of a private research Discord server and several close friends.

Intellectual & Cultural Interests

The capture data reveals an unusually wide range of intellectual engagement, often pursued in rapid succession within single browsing sessions. See dedicated articles: Music & Listening, Film & Media, Intellectual Life.

Key Interest Clusters

Philosophy of Guardian

Oliver approaches Guardian through a Heideggerian framework, organizing the project's conceptual documents under titles like "The Heidegger Frame," "Guardian Angel," "The Gwern Frame," and "The Business Model (TSMC / Costco)."[19]

The core question, as her documents frame it: what does it mean to build something that truly witnesses a life? The system distinguishes between "grounded" and "haunted" memory—grounded memory is traceable to evidence, while haunted memory fills gaps with inference. Guardian aspires to the former.[20]

In an extended Discord conversation in a research community on March 14, Oliver discussed training models on personal life data with several other researchers. The conversation explored RAG for episodic memory, finetuning for self-model, and provenance as a grounding mechanism.[21]

Notable Quotes

Paul Han

Cofounder of Haptica · Most-referenced person in Guardian capture data

Paul Han is a cofounder of Haptica and Jackie Oliver's closest collaborator. He appears in the Guardian capture data more frequently than any other individual, with references spanning audio transcripts, Discord direct messages, desktop screenpipe, browser activity, and visual captures across the majority of recorded days.[22]

Relationship with Jackie

The capture data paints a picture of two cofounders in close daily contact. Paul and Jackie communicate primarily via Discord DMs[23], in person during evening work sessions, and occasionally by phone. They discuss everything from product strategy and hardware reverse-engineering to Reddit threads, AI relationships, and DJ Khaled compilations.

On April 5, Paul and Jackie had a grounded conversation about fundraising strategy, during which Paul suggested finding investors "who have some distance. They like the tech in the way they like the Oura Ring."[24] Jackie noted that Paul was "not negative" in his approach—measured and constructive.

On April 6, a more direct exchange surfaced in the capture data, with Jackie and Paul debating the best product strategy—the kind of frank cofounder conversation that happens when both people care about getting it right.[25]

Technical Contributions

Paul's technical work includes camera hardware research and infrastructure work.[26] He and Jackie share a local network for development.[27]

Haptica

Robotics training data company · Founded by Jackie Oliver and Paul Han

Haptica is a technology startup building haptic gloves and wearable cameras for capturing physical robotics training data. The company's thesis is that real-world demonstration data—captured from humans performing physical tasks—is the critical missing ingredient for training robot foundation models, and that simulation alone cannot close the sim-to-real gap.

Product Vision

Haptica's product captures the full sensory experience of human manipulation: force feedback, pressure sensing, proprioception, and visual context. This data is intended to train imitation learning systems (behavior cloning, action chunking, diffusion policies) that can transfer to physical robots.

Competitive Landscape

Oliver conducted extensive competitive research on March 14, 2026, compiling a guardian-competitor-notes.md with assessments of the wearable AI landscape. Companies evaluated include:

Fundraising

As of April 2026, Haptica is seeking angel investment. Oliver and Paul Han have been actively networking at Bay Area tech and builder events.[30]

Guardian (Project)

Continuous life-capture and memory system · Active development March–April 2026

Guardian is a continuous life-recording and memory-synthesis system built by Jackie Oliver. It captures data from four sources (desktop screenpipe, GoPro bodycam, L810 audio recorder, and Android phone), processes raw capture into structured memory chunks, and provides evidence-backed retrieval through a read-only MCP (Model Context Protocol) server.

The Capture System

Source Hierarchy

Guardian observes a strict source priority: (1) L810 recorder audio is the canonical speech source; (2) Desktop screenpipe provides work activity, OCR, and app usage; (3) GoPro provides visual/environmental context but is NOT reliable for speech ASR; (4) Phone capture provides screen activity and notifications.[32]

Clock Synchronization

Cross-device pairing requires empirically measured clock offsets between capture devices, applied during the alignment stage of the pipeline.[33]

Data Statistics

Memory Architecture

Guardian processes raw capture through a multi-stage pipeline into structured memory chunks. Chunk types include:

• time_block — bounded periods of activity forming the day's timeline
• content_transcript — speech from audio sources
• content_visual_summary — scene descriptions from GoPro video
• content_activity_summary — desktop/app activity from screenpipe
• biography — narrative day summaries synthesized from all sources
• reflection — interpretive observations about patterns
• claim — specific factual assertions with evidence chains

Philosophical Foundation

Guardian's design principles are explicitly philosophical. Oliver's internal documents frame the project through Heidegger: it is a witness, not merely a recorder. The system distinguishes "grounded" memory (traceable to evidence) from "haunted" memory (inference without provenance). Every claim must be traceable to a memory chunk or content span; gaps are surfaced as uncertainty rather than papered over.[34]

Key Technical Decisions

• The system is read-only by design. The memory retrieval layer never writes or mutates the database.
• Capture gaps are not interpreted as sleep. Only when ALL sources are dark for 4+ hours is sleep inferred.
• Desktop audio is frequently media playback (music, podcasts, YouTube), not live speech, and must be tagged accordingly.

Music & Listening

Audio culture in the Guardian capture data

Music is a constant presence in the Guardian data, functioning as background texture during work sessions, emotional signaling during drives, and an active subject of research during browsing sessions. Oliver listens primarily through YouTube Music on the Linux desktop, with audio routed through a FiiO USB DAC.

Documented Artists & Works

The pattern suggests broad musical taste spanning shoegaze, French chanson, classic rock, hip-hop, electronic, tango, and Japanese jazz-fusion—with the specific choice often keyed to the activity (driving alone = contemplative/atmospheric; working = electronic/rock; browsing = whatever the rabbit hole surfaces).

Film & Media

Cinematic references in the Guardian data

Oliver's film interests, as captured through browser activity and YouTube viewing, skew toward art cinema and directors whose work deals with memory, time, and the visual representation of interior experience—themes directly relevant to Guardian's design philosophy.

Directors & Films Referenced

A notable pattern: Oliver doesn't just watch films; she investigates them. The Porco Rosso session spawned research into Howard Hughes, Chuck Yeager, Miyazaki's relationship with American archetypes, and the Crunchyroll News archive. The Pharrell session led to questions about corporate sanitization and music copyright law.[36]

Intellectual Life

Research patterns, reading, and curiosity threads

The capture data reveals a person who treats the internet as an instrument of investigation rather than passive consumption. Single triggers—a tweet, a video, a search result—regularly spawn multi-source research sessions lasting 30–60 minutes, jumping between Claude, ChatGPT, Wikipedia, YouTube, X, and specialized sources.[37]

Research Deep Dives (Selected)

Regular Reading

Marginal Revolution and Matt Levine appear as regular reads.[38] Gwern.net is read deeply and discussed with Claude. The New York Times and Wall Street Journal appear in browsing activity. Oliver maintains tabs across multiple Claude, ChatGPT, and Grok sessions simultaneously, using different AI tools for different query types.

The Keoni Meeting

A pivotal pitch meeting referenced across multiple capture days

The Keoni Meeting refers to a pitch or business meeting between Jackie Oliver, Paul Han, and an individual identified as Keoni. It is one of the most-referenced events in the Guardian data, with follow-up activity spanning multiple days in early April 2026.[39]

The Meeting

The meeting itself appears to have taken place in the days before April 3, 2026. A browser research session on April 3 references "a pitch meeting" with several attendees.[40]

Aftermath and Speaker Attribution

Significant effort was spent in the days following the meeting correcting speaker diarization errors. The Guardian pipeline had misattributed certain statements—the capture data shows Jackie and Paul discussing "specific attributions of statements made by Paul, Keoni, and [Jackie]" on April 6.[41] A terminal session displayed a document titled "The Keoni Meeting - Full Story."[42]

Key attributions from the meeting (from existing memory notes): Paul said "fire" (in some context), Paul got stuck on privacy concerns, Jackie rescued the conversation, and Keoni cited the Oura Ring as a reference point.[43]

Post-Meeting Activity

Between April 5–6, Jackie spent extensive time querying Guardian's memory system for information about Keoni.[44] Speaker identification for meeting participants proved difficult due to missing voice enrollment data.[45]

Gwern Conversation

March 14, 2026 — Finetuning, dynamic evaluation, and AI poetry

On March 14, 2026, Jackie Oliver had a chat conversation with Gwern (the pseudonymous writer and researcher known for gwern.net) about finetuning, dynamic evaluation, AI poetry, and LoRA vs full finetuning tradeoffs.[46]

The conversation occurred around 07:14 UTC and was brief, but its ideas reverberated through Jackie's subsequent work on Guardian's conceptual framework. Gwern's advice on Haptica's business model and his "thinking machines" concept influenced Jackie's approach to personalized AI philosophy.[47]

Later Gwern Reading

On March 23, Jackie spent extended time on Claude discussing "Gwern story about isolation and irrelevance."[48] On March 29–30, she read Gwern's essay "Why Tool AIs Want to Be Agent AIs" deeply, with multiple return visits to the gwern.net page and cross-referencing Twitter discourse (Tenobrus: "@shlevy gwern laid out many...").[49]

San Francisco

Primary location in Guardian capture data

The San Francisco Bay Area is Jackie Oliver's home base and the setting for nearly all Guardian capture data. Specific locations referenced include:

Neighborhoods & Locations

• Mission District — Google Maps exploration of restaurants and neighborhood research[50]
• Various SF neighborhoods — SoMa, FiDi, Mission Dolores area appear in browsing and errands[51]
• Tech & cultural venues — Various music venues, networking events, and meetups researched[52]

Events Researched

Oliver actively researches Bay Area tech, AI, and social events using ChatGPT, Google, and Luma. The capture data shows regular browsing of founder meetups, VC events, and electronic music listings.[53]

Work Patterns

Observable patterns from 34 days of continuous capture

Tools & Technology

Observable Behaviors

• Rapid context-switching: Rarely more than 10–15 minutes on a single task before branching to a new thread.
• Multi-AI workflow: Claude Code, Claude web, ChatGPT, and Gemini used simultaneously, with different tools for different query types.
• Rabbit hole architecture: Single triggers spawn multi-source investigations lasting 30–60 minutes.
• Interleaved building: Product work surfaces in brief terminal bursts between browsing sessions, not in dedicated blocks.
• Video speed adjustment: Documentaries at 1.5–2x speed; music videos at normal speed.
• Music as constant companion: Always playing, choice keyed to activity and mood.

Other People

Recurring individuals in the Guardian capture data

Data Coverage

What Guardian has recorded and what's missing

Sources & Methods

This wiki is compiled from the Guardian memory database, a snapshot from April 2026. Sources include:

• Memory chunks: time blocks, biographies, transcripts, visual summaries, activity summaries, reflections, claims
• Content spans: Raw evidence from cameras, audio recorders, desktop capture, and AI interpretation
• Daily biographies: 26 narrative summaries synthesized across all modalities
• Structured metadata: Activities, people, locations, decisions, emotional signals

Coverage Map

Known Limitations

• Capture gaps ≠ inactivity. A gap may be a dead battery, deliberate pause, or device failure. Only infer sleep when all 4 sources are dark for 4+ hours.[60]
• Desktop audio contains media. Music, podcasts, and YouTube audio are captured alongside live speech. Not all transcripts represent Jackie's own words.
• Speaker identification is unreliable. Automated speaker diarization requires strict confidence thresholds. Most speakers beyond Jackie remain anonymous clusters.
• Android phone capture is scaffolding-stage. Phone activity is the least complete modality.
• GoPro audio is not canonical for speech. The recorder (L810) is the trusted speech source.