People forget promises, names, appointments and half-heard instructions. A new device class — AI pendants, clip recorders, wristbands and camera glasses — captures conversations and daily context, transcribes them, and builds a searchable personal memory with automatic reminders and to-dos. The technology works; the hard part in the EU is law and etiquette: recording other people who never agreed to it.
Problem & value
- Today: memory relies on notes typed in the moment (KNOW-018), phone voice memos, or scheduled meeting recorders (SPEECH-001). Spontaneous conversations (doctor, landlord, school teacher, family planning, a street-corner idea) are lost or remembered wrongly.
- AI value: low-friction capture, retrieval by natural-language question, automatic extraction of commitments and dates, and context across weeks. Strong accessibility benefit for people with memory impairments or ADHD (to be validated per person; no clinical claims here).
- Market signal: consolidation shows the category is strategic. Amazon acquired Bee (July 2025) and showed it at CES 2026; Meta announced the acquisition of Limitless in December 2025, stopped selling the pendant and ended service in several regions including the EU and UK [1]. Plaud launched the clip-on NotePin S at CES 2026 [1]. Lesson: the cloud service behind a gadget can vanish, so data portability matters.
- No credible peer-reviewed figures on productivity gain exist; treat vendor claims as marketing.
Variants
- Audio pendant / clip / wristband (Plaud NotePin, Bee, formerly Limitless): always-on or tap-to-record, phone app does summarisation.
- Smart glasses with camera and microphones (Meta Ray-Ban / Ray-Ban Display, Even Realities G1/G2, Brilliant Labs Halo, XREAL/Rokid-class devices): add "what did I just see / where did I park".
- Phone-only lifelogging: background recording or screen/notification journaling, no extra hardware.
- Assistive memory aid: reminders for people with mild memory problems, often with a carer.
- Private local variant: all processing on a home server (PERS-005).
- Out of scope here: professional meetings (SPEECH-001), photo libraries (PERS-006), blind-user scene description (PERS-007, CONV-027).
How it works
- wearable (mic / camera, button or wake-gesture, VAD)
- BLE/Wi-Fi sync to phone app (encrypted)
- voice-activity detection + ASR + speaker diarisation (optionally on-device)
- [glasses] keyframe sampling
- VLM captioning
- LLM: summary, topics, commitments, entities (people, places, dates)
- embeddings
- vector index + timeline store
- reminders / to-dos
- calendar & task apps (AGENT-012)
- retrieval: "what did X promise me?" (RAG over own memory, with source snippets)
- Capture policy is the key design element: tap-to-record or visible indicator, geofenced "never record" zones (bathroom, doctor, school), automatic pause on detected third-party speech in private settings.
- ASR + diarisation: Whisper-class models, plus diarisation; speaker naming needs voice enrolment, which touches biometrics (see Risks).
- Memory layer: hierarchical summaries (day, week) plus raw snippets for citation, to limit hallucination.
- Human-in-the-loop: the owner confirms extracted to-dos, can delete any segment, and can mark recordings "do not index".
Data preparation
The "data" is mostly created by the device itself: audio, photos, calendar entries, contacts and location. It lives on the wearable, in the phone app and in the vendor cloud; paper or ERP systems play no role. Typical gaps are noisy audio, overlapping speakers, missing speaker names, duplicate recordings, no consistent timestamps or place labels, and above all no consent of the people being recorded. Preparation steps: (1) decide what is recorded at all (push-to-record, no always-on in private rooms); (2) enrol a few known voices for diarisation and add a contact list so names resolve; (3) transcribe locally (Whisper, faster-whisper) and discard raw audio after a short retention period; (4) chunk transcripts by conversation, add date, place and people as metadata, and embed them in a local vector store; (5) pseudonymise third parties where possible (PREP-022). A glossary of personal terms and names helps recognition (PREP-015). Rough share of effort: about 10-15 percent, because capture is automatic; most effort goes into privacy design and retention rules. Classification of what may go into which tool is covered in PREP-024, and the data-readiness basics in GUIDE-003.
Implementation paths
Which path for whom
- Private: Path C. Conversations of third parties and daily routines are highly sensitive, and Whisper-class transcription plus a small local model runs on a phone or laptop (HW1); ready-made pendants (Path A) are convenient but send audio to vendor clouds, often outside the EU.
- Developer: Path B. Building a memory product needs fast iteration on EU-hosted speech and LLM APIs, with local processing added later once the privacy design is settled.
Path A — Buy (convenient, but cloud-bound)
- Audio wearables: Plaud NotePin / NotePin S and Note Pro (recorder with app and subscription), Amazon Bee (wrist, Alexa+ integration; check EU availability), Limitless (discontinued — do not buy).
- Glasses: Meta Ray-Ban line with Meta AI; Even Realities G-series (display glasses, audio focus); Brilliant Labs (open-source-friendly glasses).
- Software only: phone apps with background transcription and memory search.
- Check: EU data residency, DPA/privacy policy, export format, deletion on request, what happens if the vendor is acquired or closes, recording indicator for bystanders. No dominant EU hardware vendor yet; EU-hosted transcription backends exist, so ask where audio is processed.
- Choose A when: you want it working in days and accept a subscription and cloud processing.
Path B — Build on cloud APIs
- Transcription: Mistral Voxtral (EU provider), OpenAI transcription models, Google Speech-to-Text, Azure AI Speech (EU regions), Deepgram or AssemblyAI with diarisation.
- Reasoning/summaries: Claude, Gemini, GPT or Mistral models via EU regions where offered.
- Orchestration: phone app or n8n; embeddings plus pgvector or Qdrant (EU-based vendor).
- Choose B when: you are a developer building a custom memory app on generic hardware (e.g. a Bluetooth recorder or an ESP32/open pendant design).
Path C — Self-host (open source)
- ASR: Whisper large-v3 / faster-whisper, whisper.cpp for on-phone use; NVIDIA Parakeet for speed.
- Diarisation: pyannote.audio.
- LLM: Qwen, Llama, Gemma or Mistral small models via Ollama / llama.cpp; VLMs such as Qwen-VL for glasses frames.
- Memory: Postgres + pgvector or Qdrant; open-source memory frameworks (e.g. Mem0) can be evaluated.
- Hardware: HW1 suffices for ASR and a small LLM at night-batch speed (Mac mini, mini-PC); HW2 for responsive queries. All audio stays at home (PERS-005).
Path D — Custom training
Rarely worth it for individuals. Possible: fine-tuned small ASR for dialect or accented speech, or a wake-gesture classifier. Needs consent-based, well-labelled own data.
State of the art (2026)
- Transcription of clean close-talk speech is very accurate across major EU languages; noise, overlapping speakers, dialects and code-switching remain the main error sources (reviewers also report weak pendant microphones in noisy rooms) [1].
- LLM summarisation is good but can invent commitments or misattribute speakers; always keep the snippet link.
- Always-listening edge inference is limited by battery; most devices still offload to phone/cloud.
- Hardware churn is high; business models shifted from one-off purchases to subscriptions.
- Limits: pendants do not understand context outside the audio; glasses cameras raise stronger social and legal pushback than audio alone.
Shared resources & combinations
- Same stack (asr + diarization + llm-general): SPEECH-001 meeting notes, SPEECH-003 translation, SPEECH-017 hearables.
- Memory/RAG layer: KNOW-018 second brain, PERS-006 photo search — one shared personal index.
- Home server: PERS-005 provides the HW1/HW2 local backend.
- Actions: AGENT-012 calendar agent; PERS-009 family organisation.
- Visual: CONV-027 live visual assistance; PERS-007 assistive vision.
- Rights: PERS-016 for access/erasure requests towards vendors.
Effort, cost & KPIs
- Path A: minutes to days; hardware roughly 50–200 EUR plus often a monthly subscription (prices from reviews, vary) [1].
- Path B: a few weeks for a prototype; cost driven by audio hours transcribed.
- Path C: a weekend to set up; ongoing upkeep and one-off home hardware.
- KPIs: retrieval success rate on test questions, share of extracted to-dos accepted by the user, transcription word error rate, battery life, number of unintended recordings, time to delete data.
People & change
This is mainly a private use case, so the "stakeholders" are the user and the people around them. Involve family, friends and colleagues before wearing a recorder: tell them when it records, ask for consent, and switch it off on request. Recording confidential spoken words of others can be a criminal offence in Germany (section 201 StGB), so clear rules matter more than technology. Works council: no, as long as the device is private; if an employer or a developer team tests it with staff, it becomes a system that can monitor behaviour and a works council (or staff council) is involved under section 87(1) no. 6 BetrVG, and the information duty of section 90(1) no. 3 applies (see GUIDE-004). Developers offering the product need a DPO, a DPIA (systematic recording, new technology) and processor agreements (GUIDE-005). AI literacy: Art. 4 AI Act asks providers and deployers to support the AI literacy of staff, so provide short onboarding on limits such as wrong attributions and hallucinated summaries (GUIDE-006). Pilot with a few informed volunteers for two weeks. Typical concerns are surveillance and trust: address them with visible recording indicators, local storage, deletion on request and humans reviewing every summary before it triggers an action.
Risks, regulation & compliance
- Criminal law (Germany): secretly recording the non-publicly spoken word of others is an offence under §201 StGB (Verletzung der Vertraulichkeit des Wortes); using or passing on such recordings is also punishable. Consent of all participants is the safe rule. Other EU states have comparable rules [2].
- GDPR: the household exemption (Art. 2(2)(c)) is narrow; case law (CJEU Ryneš, C-212/13) limits it where recording extends to public space. Employers and professionals recording clients or colleagues are fully controllers. Vendors are processors/controllers: DPA, transfer safeguards, retention, and deletion rights apply [3]. Voice enrolment for speaker identification is biometric data (Art. 9 risk).
- EU AI Act: typically limited/minimal risk for private use; transparency duties (Art. 50) apply to the provider for AI systems interacting with persons or generating synthetic content. Art. 5 warning: emotion inference from voice or face in the workplace or education is prohibited, and untargeted scraping of facial images or real-time remote biometric identification in public spaces are prohibited or heavily restricted. A lifelogging product must not offer "mood detection of colleagues" or face recognition of bystanders. Treat workplace pendants as an employment-law and works-council topic [4].
- Special settings: medical, legal, school and confidential business settings (professional secrecy, §203 StGB) — do not record without explicit permission.
- Other risks: breaches expose highly intimate data; vendor shutdown or acquisition; coercive use within families or relationships.
- Mitigations: visible indicator and verbal announcement, consent templates, no-record zones, on-device or EU-only processing, short raw-audio retention (summaries only), encryption, exportability.
Getting started
- Decide the use: personal reminders, accessibility, or work. Exclude work and confidential settings first.
- Test tap-to-record with a clip recorder or phone app for two weeks; always announce recording.
- Check the vendor DPA, EU processing, export and deletion; avoid products whose service is ending.
- Define no-record zones and a 30-day raw-audio retention.
- Optional: move to a local stack (Whisper + pyannote + small LLM) on a home server.
Sources
- The Gadgeteer — Best AI wearables 2026 (Bee, Plaud, Limitless context) — https://the-gadgeteer.com/2026/05/05/best-ai-wearables-2026/
- §201 StGB, Verletzung der Vertraulichkeit des Wortes — https://www.gesetze-im-internet.de/stgb/__201.html
- Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Regulation (EU) 2024/1689 (AI Act), Art. 5 and 50 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- CJEU C-212/13 Ryneš — https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62013CJ0212
- pyannote.audio — https://github.com/pyannote/pyannote-audio
- whisper.cpp — https://github.com/ggml-org/whisper.cpp