Under a storage limit, which transcripts should an on-device assistant keep, and what survives?

What it found

Seven simulated days of 1500 utterances each with 6 planted facts per day. BM25 finds every fact (hit@5 = 1.00) when the key words are rare, which is the easy case. With the store limited to 10% of the week, dropping oldest or random utterances keeps 0.11 and 0.10 of the facts; keeping the highest mean-IDF utterances keeps 0.93, because the facts are rare-word lines by construction. An importance-ranked extractive summary at 5% of a day holds 85% of its facts, against 3% for the first lines. Index build 0.04 s, query 1.9 ms on one container CPU. Stylised: synthetic words, lexical queries, no paraphrase, no speech errors. Produced with AI assistance.

A figure from its own output

Paper, code and results

• README and method
• Experiments and results
• Code
• Tests

More in Ambient AI assistant

• Which parts of an always-on Android listening service can be tested without a phone?Testable Kotlin core of an always-on listening service (re-framing, energy VAD, utterance segmenter, headset-drop policy; 7 tests pass under a shim) plus an unbuilt Android foreground-service sketch; no SDK, no device numbers
• When is a glance better than speech for an always-on assistant?When a glanceable display beats speech for an always-on assistant: a stylised utility model over noise, company, free eyes and message length, with sensitivity and twin-gap checks
• Does urgency, and a chime-plus-glance mode, change how an always-on assistant should deliver a message?Urgency classes and a chime-plus-glance mode added to the glance-vs-speech utility model for an always-on assistant: a stylised model with sensitivity and twin-gap checks
• How much reply latency can an always-on assistant hide by starting its pipeline before the speaker has finished?How much reply latency starting the ASR/LLM/TTS chain during the end-of-speech hangover hides, from the measured ambient-pipeline stage times and an assumed within-turn pause model (hold-until-endpoint vs eager; extra runs per turn)
• Can a sentence-completeness cue let an always-on assistant stop waiting so long before it replies?Can a sentence-completeness cue let an always-on assistant stop waiting so long before it replies: a stylised two-hangover endpointer cuts the mean wait from 913 to 574 ms at a 10% cut-off budget if no pause sits at a complete point, but only to 837 ms if half do; cue and pauses assumed.
• When is a second or third earbud tone worth it before a glance?How many repeated earbud tones before a glance are worth their annoyance, by urgency, tone gap and masking, in the ambient-urgency utility model (stylised)

All 46 Ambient AI assistant projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.