The models log uses

log uses six models. Two run on your own machine, with no key and no internet once they are downloaded. Four are hosted services you add a key for. This page says what each one does, what your machine needs, and how the indexes they read are built.

At a glance

ModelRunsWhat it does in logSizeKey
all-MiniLM-L6-v2 (8-bit)your machineturns text into meaning vectors: log search, your voice index, the knowledge base, matching what you typed to a play, and (with char n-grams) the voice gate's AI-conventions detector trained on you23 MBnone
roberta-base-go_emotions (8-bit)your machinethe top three emotions of each new message someone sends you, a feature of who they are and where they are (pipeline labels)125 MBnone
LFM2.5-VL-450M (4-bit)your machinedescribes every picture, gif and video frame someone sends you, as text352 MBnone
SubQ subq-snap-1.1hostedthe voice gate (does a draft sound like you), reading an unknown page state, briefs, meme tasten/aSUBQ_API_KEY
Jina jina-embeddings-v5-omni-small and jina-reranker-v3.5hostedranking a chat's history for the voice gate, and the old AI tells index (read only with LOG_AI_TELLS=jina)n/aJINA_API_KEY
AssemblyAIhostedwhat was said in a voice note or a videon/aASSEMBLYAI_API_KEY
Inworld inworld-tts-2 (or Fish Audio s2.1-pro)hostedspeaks a voice note in your cloned voice, only when you askn/aINWORLD_API_KEY or FISH_API_KEY

The model names are in the code, so this table can be checked against it: EMBED_MODEL (src/search/conversations/types.ts), VISION_MODEL (src/media/describer/types.ts), VOICE_MODEL (src/twitter/copy/voice/types.ts), JINA_EMBED_MODEL and JINA_RERANK_MODEL (src/voice/aiTells/types.ts), and the voice providers in src/extensions/inworld and src/extensions/fishAudio.

Where each one sits

Without the hosted keys, log still reads, searches and describes pictures. The voice gate, transcripts and voice notes need their keys.

What your machine needs

Measured on this project's own machine on 2026-11-09 (Windows 11, 16 threads, CPU only, no GPU):

LoadEach useMemory
MiniLM0.17 s2.8 ms p50, 9.8 ms p95 to embed one message (20 runs)small, inside the process using it
go_emotions RoBERTa1.3 s24 ms p50, 35 ms p95 per message on a quiet machine; run after the who and where labels so it never holds the pushinside the push process
LFM2.5-VL5.7 s6.3 to 7.4 s per picture (3 runs)peaked at 1.25 GB
The describer idle, between picturesn/an/aabout 100 MB

So, at least:

  • A 64-bit machine. Both local models run on the CPU, so no GPU is needed.
  • About 2 GB of free memory for the picture model while it works. It stops itself, loudly, past 4 GB (MEMORY_CEILING_MB), and only one describer runs per machine.
  • About 400 MB of disk for the two models. They download once, the first time they are needed, into node_modules/@huggingface/transformers/.cache.
  • ffmpeg and ffprobe on your PATH, for videos and voice notes (LOG_FFMPEG and LOG_FFPROBE point elsewhere).
  • Bun and Python 3.12. The installer sets both up (install). log voice train also needs scikit-learn in that Python (uv pip install --python <LOG_PYTHON> "scikit-learn>=1.5,<2"); without it the trainer runs through uv run --with scikit-learn.

A picture takes seconds to describe, so pictures queue (at most 20) and are described one at a time. Until one is described it shows as [picture … not described yet], so nobody answers it blind.

How the indexes are built

Every index is built from what is already in log.db or what log has saved, and each one is rebuilt in full by its command. Run them in this order on a new machine:

  1. log search index reads every conversation once through each platform's own session and builds two indexes in ~/.logCli/search: one by keyword (BM25) and one by meaning (MiniLM vectors, searched with HNSW). After that, each new read adds only what changed. log search "<query>" uses both and merges the results.
  2. log voice index builds your voice from the conversations step 1 saved: each of your replies paired with what was said just before it, embedded with MiniLM, plus a profile per chat, person and platform. It reads nothing new.
  3. log voice train builds the voice gate's AI-conventions detector: your own messages in log.db against the drafts agents wrote for your chats (read from the agent's sessions and from every draft the gate grades, kept in log.db), plus, with --fresh <n>, drafts from several frontier models for your real past moments, each beside what you really said (it prints the cost first and stops at --spend, $3 by default). It keeps the words and habits models use and you don't, each beside your nearest real phrasing, and a small classifier head (char n-grams plus MiniLM, fitted with scikit-learn in LOG_PYTHON, applied in TypeScript), with one line per message length so 95% of your own held-out messages pass. The voice gate starts it in the background when it is missing or older than step 2. log voice aiIndex builds the old Jina phrase index, read only with LOG_AI_TELLS=jina.
  4. log kb sync (optional) puts the rules, skills, docs and agent memory into the knowledge base on Redis, searched the same two ways.

Pictures and videos are not indexed by a command. The describer (log media describer, started by the inbox hook) describes each one as it arrives, and its text is searchable like any message. log media status shows the queue.

The voice gate's embeddings are made once and kept in log.db by the text's hash and model, so each send only embeds what is new. An index made with a different model is rebuilt.

Why these models

  • Small enough to run on any machine. MiniLM and LFM2.5-VL run on a CPU, so reading, searching and describing never need the internet or a key.
  • Text for every model. Every picture, voice note and video becomes text when it arrives (how it works), so an agent whose model can't see images still knows what was sent.
  • Hosted only where it pays. The voice gate and transcripts are worth a hosted model, and each one's balance is in log balance.