v0.4.0-beta.1 · local-only

A local memory palace. For everything you scattered.

SuperTon ingests your notes, PDFs, and AI-chat transcripts into a verbatim store, then a tiny local model answers your questions — grounded only in what you've fed it. Nothing leaves your laptop.

Two commands. No cloud account. No telemetry.

01Install SuperTon
$curl -fsSL https://raw.githubusercontent.com/therahul-yo/Superton/main/install.sh | shcopy
Requires Python 3.11+. The installer pulls uv if missing, then installs SuperTon as a uv tool. Alternative: uv tool install superton from PyPI, or clone the repo and run uv run superton.
02Initialize the palace
$superton initcopy
Pulls Ollama if missing, downloads MiniCPM5 (~700 MB), builds the superton model. Arrow-key theme picker included.
+Optional · install Ollama yourself
$brew install ollamacopy

Ingest. Ask. Cite.

~/superton · superton

What's already in the box.

I.

Multi-source ingest.

PDFs, markdown, code, and AI-chat transcripts from Claude Code, ChatGPT, Cursor, and Amp. One add command, one verbatim store.

II.

Semantic + lexical search.

Hybrid retrieval via MemPalace with a SQLite FTS fallback. Every answer footers with the drawer IDs it cited — refusal over confabulation when the palace is silent.

III.

MCP-ready out of the box.

superton mcp serve exposes 29 tools over stdio so Claude Code, Cursor, and Gemini CLI can read and write your palace directly — no copy-paste round-trips.

IV.

One tiny model.

openbmb/minicpm5 — 1B params, ~700 MB download, 128K context. Fast first token on any laptop; the palace provides exact recall, the model only phrases the answer. Override with SUPERTON_BASE_MODEL.

Fast enough to forget the spinner.

Measured on Apple M4 MacBook Air (16 GB) with 2,000 synthetic drawers. Re-run via scripts/bench.py.

~/superton · bench
# benchmark — apple m4 macbook air, 16 gb CLI cold start (superton --version) ~87 ms Ingest throughput ~11,500 drawers/sec Retrieval latency p50 / p95 0.99 ms / 1.10 ms Storage per drawer ~840 bytes (SQLite + FTS) # first-token latency depends on your ollama setup and base # model — typically ~2 s on M4 with openbmb/minicpm5.

Three opinions, kept honest.

I.

Verbatim beats summarized.

Drawers store the original text. A summary loses the half-remembered phrase you'll actually search for in six months.

II.

Refuse over confabulate.

When the palace is silent, Superton says so — and suggests the closest files. No invented dates, names, or numbers.

III.

Local until proven otherwise.

No API keys. No telemetry. No SaaS dependency. The palace lives on your laptop; the model lives in your Ollama.

drawers in, nothing forgotten.
github · docs · apache-2.0