Private · internal plan · do not share
← Game plan  ·  Memsist
Replacement Roadmap

Off Limitless, onto our own stack — before 2026 ends.

Our lifelog runs on the Limitless Pendant. Limitless is being shut down. This is the plan to own the whole pipeline ourselves — capture, transcription, and recall — with no monthly cloud and no per-token AI billing.

Built 2026-06-29 · research-backed · supersedes ad-hoc notes
The clock is real and external.

Meta acquired Limitless on December 5, 2025. Hardware sales stopped immediately; existing Pendant owners were moved to a free plan with roughly one year of support — so the cloud and the API are expected to go dark around late 2026. (The companion Rewind product was already killed Dec 19, 2025.) The Limitless API — including /v1/download-audio for raw Ogg/Opus — is still live today. Every week we wait is data we may not be able to pull later.

01The shape of the replacement

Three layers, each independently ownable. The principle is the same as the rest of the network: local-first, you own the data, no per-token API spend. The only outside accounts are cheap storage (Backblaze B2) and a free HuggingFace token to download speaker-recognition weights.

1

Capture — a pendant we control

An always-on wearable that records the day and hands us the raw audio, not just somebody’s cloud summary. The winner is Omi: open-source hardware and firmware, self-hostable backend.

DECISION: Omi dev kit now · phone + Auphonic as the zero-lock-in fallback
2

Transcribe — on a machine in the house

A home PC with a consumer GPU runs WhisperX (large-v3) for transcription plus pyannote for speaker labels (who said what). The droplet can’t do this at quality, so it instead runs a fast “live today” pass on the newest audio (whisper.cpp) and the GPU PC heals it to the polished version when it’s awake.

DECISION: GPU PC = engine · droplet = always-on front door + live preview
3

Recall — ask anything, answered locally

The droplet keeps everything in SQLite + full-text search, rolled up into day / week / month / quarter. Questions are answered with local embeddings + keyword search (sqlite-vec + FTS5) returning dated, cited snippets — with an optional small local model for fluent paragraphs. No Anthropic/OpenAI billing, ever.

DECISION: replace today’s billed /ask with the local hybrid retriever
  OMI PENDANT ──▶ Backblaze B2 (raw audio, we own it)
       │                  │
       │     ┌────────────┴───────────────┐
       │     ▼                            ▼
       │  DROPLET (always on)        HOME GPU PC (the engine)
       │  • whisper.cpp "live today" • WhisperX large-v3 + pyannote
       │  • SQLite + FTS5 search     • writes speaker-labeled JSON
       │  • day/week/month/quarter   • POSTs "final" → droplet
       │  • local Q&A (sqlite-vec)        (heals the live preview)
       ▼
  lifelog.walhus.com  ◀── ask anything, fresh to the minute

02Already shipped (today)

The recall layer is live now on lifelog.walhus.com

03The device decision

Most “AI pendants” being hyped right now fail our one non-negotiable: give us the raw audio. Here is where each lands.

DeviceAlways-onRaw audio?Lock-inVerdict
Omi (open-source)YesYes — self-hostNone (open)Buy now
Phone + Auphonic recorderYesYes (local files)NoneFree fallback / insurance
Plaud Note / NotePinNo — push-to-recordYes (export)Cloud transcribeWrong capture model
Limitless (incumbent)YesYes — for nowShutting downStopgap only
Bee (Amazon)YesNo — deleted after transcribeAmazon-lockedAvoid
Compass / FieldyYesNo export APIClosedAvoid

04Timeline to year-end

This week — do not waitBack-pull the entire Limitless archive

Run a full export of every lifelog transcript and the raw audio via /v1/download-audio to Backblaze B2 while the API is still up. This is the irreplaceable step — the history can’t be re-recorded.

JulyOrder an Omi dev kit · stand up the transcription pilot

Run Omi in parallel with Limitless. Wire the home GPU PC: pull a chunk from B2 → WhisperX + diarization → POST to the droplet. Benchmark the real throughput on our card.

August–SeptemberCut the recall layer off paid API

Replace the billed /ask with local embeddings (sqlite-vec) + FTS5 hybrid search and dated-snippet answers. Add the “live today” droplet quick-pass so the log is current without the GPU PC awake.

October–NovemberSwitch primary capture to Omi

Once the pipeline is proven, Omi becomes the daily driver; Limitless drops to backup until its cloud dies. Speaker-name enrollment (“SPEAKER_01 = Dot”) and the relabel UI land here.

DecemberFully independent

Limitless can shut down with zero data loss and no feature regression. Everything — capture, transcription, recall — runs on hardware we own.

05Guardrails

06The next sense — sight

Everything above makes the lifelog independent by ear. The next milestone adds vision — and the principle is identical: capture the raw frames, keep them home, no per-token billing. The public framing lives at memsist.com/vibelens.

4

See — a wearable camera we can pull frames from

Same non-negotiable as audio: the device must hand us the raw capture, not a vendor’s cloud summary. VibeLens (a head-worn POV camera — live preview, open-ear audio, Wi-Fi transfer) is the candidate reference device, the visual analogue of Omi. If it won’t export raw frames, it’s out — same rule that ruled out Bee and Compass.

DECISION: evaluate VibeLens for raw-frame export · phone camera as the zero-lock-in fallback
5

Recognize — a vision pass on the home GPU PC

The same GPU PC that runs WhisperX runs the vision pass: sample frames on motion/scene-change, caption them locally, detect faces/objects/text (OCR on notes & labels), and write a dated, searchable visual note. It joins the existing SQLite + FTS5 spine so “where did I leave my glasses?” returns a timestamped snippet — text and, when it helps, the frame. Local embeddings only; no Anthropic/OpenAI billing.

DECISION: on-device vision (local caption + OCR + face/object) → same FTS5 store · nothing to the cloud
After audio independence (Q1 2027 target)Visual pilot on one consenting user

Once capture+recall are fully off Limitless, run VibeLens in parallel for a single consenting user (self first). Prove: raw-frame export, the local vision pass throughput on our card, storage growth per day, and the recall UX (“show me” vs “tell me”). Consent and privacy stakes are higher with a camera — treat it like the care relationship it is.

Memsist · Memory Assist · private internal roadmap · 2026-06-29 · sight layer added 2026-07-10