Narrative App & Platform

The clip camera got the headlines, but the harder engineering problem — and the actual product, in Martin's telling — was the software behind it: a cloud service that took a firehose of near-identical photos (up to 2,000 a day per user) and turned them into something a human could actually look at. The "moments" algorithm clustered the day into a handful of natural segments, scored every photo for quality, and surfaced one representative keyframe per segment on a mobile timeline — automatically, with no shutter button and no manual editing. CTO Björn Wesén called that cloud software the hardest part of building the company (blogs-wsj-com ↗), and the app built around it — free on iOS and Android, later joined by a desktop uploader and a web viewer — was the layer users actually lived in, long after the novelty of the clip itself wore off.

See also: Narrative for the company story, Narrative Clip 1 and Narrative Clip 2 for the hardware, and Memoto Kickstarter Campaign for the launch.

The moments algorithm: turning a firehose of photos into a story

Ambient color clustering, then a quality score. In his own later account, Martin described the mechanism plainly: the algorithm used ambient color histograms to cluster photos into segments of the day — "moments" — and then, within each cluster, picked a cover photo using technical image-quality metrics: color saturation, contrast, detail, and whether a face was present. ▶ 11:05

• The core innovation was this two-step process — segment the day by ambient color clustering, then assign each photo a quality score — rather than any single clever trick. kaptur-co ↗
• Roughly half of all captured photos were filtered out before a user ever saw them, specifically to avoid walls, ceilings, and floors and keep the surfaced set focused on people and variety. ▶ 11:56
• Cloud algorithms combined image-quality analysis with location metadata to group photos into discrete moments — the processing happened entirely server-side, not on the device. engadget-com ↗
• Reviewers who saw the pipeline described in detail counted the signal set as GPS, time, lighting, motion, colorfulness, photo-composition symmetry, and face detection — all feeding a per-photo quality score. cnet-com ↗, weartested-org ↗
• A "quality score" system explicitly filtered out photos judged to carry no new information, rather than simply picking one photo per time window. ▶ 1:15
• The service also handled cleanup automatically: correcting photo rotation to the horizon using onboard sensor data (even when the camera was worn at an angle), and discarding extremely dark or blurry frames outright. istartedsomething-com ↗, weartested-org ↗
• A typical day compressed down to about 20–30 keyframes standing in for the whole day's stream. ▶ 27:25, ▶ 2:07
• The algorithm's output didn't always match a user's own sense of what mattered — the moments it surfaced could diverge from what the person actually remembered as significant about their day. fastcompany-com ↗
• A known technical gap: because the algorithm processed uploads in batches, photos from an earlier upload never got retroactively re-grouped with a later one, even when they matched the same clustering criteria — producing duplicate moments instead of one merged moment. ▶ 5:55

From clip to cloud: the capture pipeline

Every generation of the app depended on getting photos off the camera and into Narrative's servers, and for most of the product's life that pipeline ran through a computer, not the cloud directly.

• The Clip 1 had no wireless radio: photos were pulled off via micro-USB into a Mac or Windows desktop app, which then uploaded them to the cloud — a proprietary piece of "uploader" software that acted as a mandatory middleman. blogs-wsj-com ↗, mobilesyrup-com ↗, digitaltrends-com ↗
• At roughly 2,000 photos captured per day, the camera auto-uploaded to cloud storage (reviewers noted Amazon-hosted infrastructure early on) the moment it was next connected and charged; server-side algorithms then ran the moment-selection pass. forbes-com ↗, onemanandhisblog-com ↗
• By the Narrative era, photos were not stored permanently on the user's own computer at all — they lived on Narrative's servers, based in Sweden, and users paid an ongoing subscription to keep them there and accessible. ▶ 5:18, ▶ 28:12
• A separate desktop companion app handled the device-to-phone handoff on some workflows, transferring photos from the Clip toward the phone app for viewing. ▶ 0:25
• Martin was candid about the cost of this: managing photo data at scale was expensive for Narrative — users captured on the order of 1,000 photos a day each — which meant complex backend infrastructure and, in his words, serious privacy concerns that came with holding that much personal imagery. ▶ 24:13
• Clip 2 finally removed the USB/desktop dependency for many users, adding built-in Wi-Fi, Bluetooth, and GPS so photos and (later) 1080p video could sync wirelessly straight to the app. dpreview-com ↗

The app: timelines, moments and manual control

"A personal automatic Instagram." That was Narrative's own framing for the app in a 2013 first-look: each day got auto-segmented into roughly ten chunks, a filter trimmed the continuous photo stream down to the most representative shot per chunk, and the design goal was that any single selected photo should tell you, at a glance, which part of the day it came from. ▶ 0:02

• Photos surfaced in the app were deliberately curated to be Instagram-friendly and visually appealing — the pitch was automatic curation, not a raw firehose. ▶ 1:38
• Moments could be shared via Instagram-like URLs, and users retained manual override: selecting a different cover photo to better represent a given moment, or sharing an individual photo to email or Facebook directly. ▶ 1:38, ▶ 2:29
• A double-tap on the camera let a wearer manually flag a moment in real time; that photo was then automatically included in the trimmed set — a way for the user to signal importance directly to the algorithm rather than hope it guessed right. slashgear-com ↗
• Depending on the release, moments were presented either as fixed ~1h45m time blocks or as "themed packages" tied to natural events (home, work, lunch, evening) — grouped using GPS and accelerometer data to cluster similar photos and hide poor-quality ones. the-gadgeteer-com ↗, mobilesyrup-com ↗
• Beyond the curated stream, the full unfiltered photo library was also browsable — reviewers noted it still contained plenty of duplicates, blur, and poor framing that the algorithm's curated view hid. ▶ 0:31
• The app supported starring favorites, scrolling through hundreds of photos quickly, and downloading photos locally — praised by at least one reviewer as genuinely impressive for the volume involved. mobilesyrup-com ↗
• Platform support broadened over time: initially iPhone and Android only, later adding a Mac/PC desktop uploader ("Narrative Uploader") and, in November 2014, a web app for desktop browsers — though the web app initially lacked the "Jump to Date" navigation feature already shipped on mobile. cnet-com ↗, thenextweb-com ↗
• No dedicated iPad app was ever built — only iPhone and Android. ▶ 3:24
• Not every review was kind to the app itself: CNET's writeup called the phone app "too pared-down to be helpful," listing it alongside poor photo framing and the paid cloud tier as the product's main drawbacks. cnet-com ↗

Sharing, search — and what didn't ship at launch

• Once processed, moments and photos could be pushed out to Facebook, Instagram, Twitter, and email directly from the app. ▶ 1:16, ▶ 1:16, digitaltrends-com ↗
• Search — filtering the photo library by time, place, or light — was part of the original vision from the Memoto Kickstarter pitch, meant to let users pull up a specific event from months of archive. kickstarter-com ↗
• It didn't ship on time: at the Narrative Clip's actual launch, search was pulled entirely because it wasn't working reliably yet, and the app shipped with chronological browsing only — a deliberate ship-imperfect decision rather than delay the hardware. slashgear-com ↗
• The platform never supported exporting a moment as a video or animated GIF — only individual still photos could be shared out, a limitation one reviewer called a missed opportunity given how naturally the clustered photos lent themselves to time-lapse playback. istartedsomething-com ↗
• Late in the product's life, as the original service wound down, Narrative planned a data-portability tool letting users download their own photos directly off company servers and pull any remaining content off physical Clip devices before support ended. engadget-com ↗

Storage, pricing and the subscription model

• Cloud storage was sold as a flat monthly fee from the beginning, with early Kickstarter backers getting the first year included for free as part of their pledge. kickstarter-com ↗
• The standard model that shipped: free cloud sync and 10GB of storage for the first year, then roughly $9/month (about $100/year) afterward to keep the moment-analysis and processing pipeline running on new uploads. istartedsomething-com ↗, cnet-com ↗, thenextweb-com ↗
• Narrative described the plan explicitly as Dropcam-style: free hardware-tier functionality, paid monthly subscription for the software that made the hardware valuable long-term. mobilesyrup-com ↗
• Clip 2 kept the same shape — 10GB free, additional storage paid — carried forward from the original Clip's pricing. ▶ 3:56

Platform ambitions: beyond the Clip itself

• By 2016, Narrative's R&D team was fielding requests for more than 1,000 potential applications of the underlying technology — well beyond what a single wearable-camera product could address — which led directly to the decision to open the platform up rather than keep it locked to one device. ▶ 11:31
• Years earlier, in a 2013 interview, Martin had already floated the same idea from the other direction: opening Narrative's cloud processing to third-party camera hardware via an API or SDK, naming GoPro and Google Glass explicitly as devices that could plug into the same moments pipeline instead of Narrative building every camera itself. slashgear-com ↗
• After the 2016 near-collapse, the reformed, employee-owned Narrative's plan was to keep charging for cloud service on the installed base while it sought a manufacturing partner to resume hardware production — the software platform, not the camera, was the asset worth preserving. techcrunch-com ↗

Worth remembering

• Privacy was structural, not a setting: everything captured landed in a private archive by default, and publishing anything required a deliberate, conscious choice each time rather than any form of automatic or default sharing. kickstarter-com ↗, ▶ 27:25
• Martin's own description of the intended payoff wasn't analytics or search — it was watching an ordinary day play back as a kind of stop-motion animation, finding meaning in the unremarkable rather than the highlight reel. huffpost-com ↗
• Even with the moments algorithm's filtering, reviewers found the ratio of total-to-curated photos ran only about 2–3:1 — meaning duplicate and near-identical shots still routinely surfaced in the "curated" timeline. slashgear-com ↗
• The original Memoto pitch imagined the lifelogging archive eventually connecting to other tracking and fitness apps, folding photo memory into a broader personal-data ecosystem — an integration that was never built out. gmanetwork-com ↗