Update wiki/always-on-ai-assistant.md
cb2e03df61be harrisonqian 2026-04-12 1 file
index 042ec74..6188305 100644
@@ -1,7 +1,13 @@
---
-status: raw
+first_captured: 2026-04-10
+sources:
+- sources/google-sheets-ideas.md
+status: explored
tags:
- ai
+- memory
+- wearable
+- context
title: always-on ai assistant
type: idea
updated: 2026-04-11
@@ -10,4 +16,8 @@ visibility: public
# always-on ai assistant
-24/7 listening, transcribing, context-aware AI companion.
\ No newline at end of file
+a 24/7 listening, transcribing, and context-aware AI companion — it is always recording your environment, converting speech to text, and building a running log of everything you say and hear. the core value proposition: you never lose a thought, a conversation, a name, a number, or a commitment again. the device closest to this vision is Limitless (formerly Rewind), but the idea is for something more ambient, lower friction, and deeply integrated with a personal AI layer.
+
+the privacy tradeoff is real and has to be confronted head-on. always-on recording is technically possible (the hardware exists — AirPods are already near your ears all day) but requires explicit design around consent, local-first processing, and strong access controls. the most compelling architecture is: local transcription on device → local embedding and indexing → only surface a query to the cloud when you explicitly ask. this keeps the raw audio and transcripts on your hardware. the stored context feeds into [[axon|axon]] for structured personal context and [[life-search|life search]] for querying what you said and heard.
+
+this is the practical near-term path for several memory ideas that are harder to build directly: the [[brain-rewinder|brain rewinder]] becomes "search your transcripts from 20 minutes ago," the [[episodic-memory-builder|episodic memory builder]] gets richer prompts from the activity log, and the [[me-model|me model]] gets training data from real conversations. it also powers [[ai-agent-reply|AI agent reply]] by giving it context about what you actually care about. the [[cluster-memory-and-context|memory and context]] cluster as a whole converges on this as infrastructure — many ideas become dramatically more achievable once you have a continuous ambient transcript.
\ No newline at end of file