Wearables

Your day, remembered, locally.

Three optional connectors ingest what your AI wearable records, turn it into clean searchable day transcripts, apply your corrections, and, under strict per-source trust gates, into memories. Everything is written to plain files under your memory directory.

Install in 60 seconds Star on GitHub

Supported devices

Each connector ships as an a-la-carte optional package. Installing @remnic/core or @remnic/cli alone never pulls one in, so you install only what you wear.

Device	Package	Native import
Limitless Pendant	`@remnic/connector-limitless`	Transcripts
Bee bracelet	`@remnic/connector-bee`	Bee facts
Omi necklace	`@remnic/connector-omi`	Omi memories

npm install -g @remnic/connector-limitless
# and/or
npm install -g @remnic/connector-bee @remnic/connector-omi

The pipeline

Every sync runs the same provider-agnostic steps per source:

pull -> off-the-record elision -> cleanup -> redaction -> corrections
     -> speaker labeling -> day-transcript store -> (optional) memories

Cleanup is deterministic and uses no LLM: it merges fragmented utterances and strips filler tokens. Redaction replaces SSN and payment-card patterns with a placeholder before anything touches disk, and you can add your own patterns. Corrections fix words the ASR keeps getting wrong. Speaker labels resolve through a persistent registry.

Transcripts are stored one markdown file per source per day at <memoryDir>/wearables/<source>/<YYYY-MM-DD>.md. They are full-text searchable but never appear as memories: the directory sits outside the memory scan roots.

Trust-gated memory creation

Wearable transcription is noisy, so a mis-transcription must not become a fact about your life. The default smart mode scores each candidate memory and routes it by trust band instead of forcing you to review every one.

memoryMode	Behavior
`smart` (default)	Automated trust pipeline. High-trust facts written active, borderline facts to the review queue, low-trust dropped.
`off`	Transcripts only. Never creates memories.
`review`	Every candidate is written pending review. Nothing enters active recall until approved.
`auto`	Deterministic gates only. Survivors written active.

Smart mode assembles a trust score from an LLM-as-judge verdict, a provenance prior per source, cross-device corroboration when another wearable covers the same day, and support from existing memories. The score maps to a decision:

trust >= 0.7   -> written ACTIVE
trust >= 0.45  -> written pending_review
below          -> dropped

Every written memory persists its evidence (trust score, judge verdict, corroborating sources) and carries provenance tags pointing back at the transcript day it came from.

CLI

remnic wearables status                     # sources, connectors, last sync
remnic wearables check limitless            # verify credentials
remnic wearables sync                       # today and yesterday, all sources
remnic wearables sync --source limitless --date 2026-06-10
remnic wearables transcript --date 2026-06-10
remnic wearables search "solar quote" --source limitless
remnic wearables memories --source limitless --date 2026-06-10

Credentials are read from environment variables such as REMNIC_LIMITLESS_API_KEY, REMNIC_BEE_API_TOKEN, and REMNIC_OMI_API_KEY. Long-lived hosts run an in-process auto-sync scheduler so today's transcript keeps growing as the device records.

MCP tools

Anything talking to Remnic over MCP gets the wearable tools:

Tool	Purpose
`remnic.wearables_status`	Source, connector, and sync status.
`remnic.wearables_sync`	Trigger a sync (source, date, days, forceMemories).
`remnic.transcript_day`	Full day transcript(s) with cross-source overlap hints.
`remnic.transcript_search`	Search stored transcripts across sources and dates.
`remnic.transcript_memories`	Memories created from transcripts, including pending-review candidates.

Full reference: docs/wearables.md.

Install in 60 seconds Star on GitHub All integrations