Memory
Salvage

lecore-bm25

Okapi BM25 + RRF with the stoplist/stemming tokenizer that makes it beat stock BM25 out of the box. Vendored from leCore (MIT, AnOversizedMooseWithSocks).

Language
Python
Pushed
Aug 2026
Size
70
License
MIT
Open repo

Who this is for

A retrieval person who wants Okapi BM25 + RRF with a stoplist/stemming tokenizer that actually beats stock BM25.

What worked

Vendored from leCore (MIT). supercontext sits next to it with NIAH / MTEB / cost-to-first-answer vs FAISS.

What failed

Easy to confuse with the whole leOS religion. This oat is a tokenizer and a bench, not an operating system.

Next three

  1. 1Publish the beat-stock-BM25 numbers in one table
  2. 2Keep leCore-the-fork separate from this SKU
  3. 3Founder who will maintain the stoplist

STACC will

Point at supercontext and the tokenizer contract.

Will not

Rebrand someone else's OS or run a hosted search API as the product.

Shipped by —, oat by STACCoverflow. No forced face. No forced token.

Same lane

Live

openzoo-ingest

Everything your machine produces, bound into your own leCore memory. Local by default; the shared brain is the only egress.

Salvage

stacc-brain

Stacc Brain — Karpathy llm-wiki at Telegram scale (Obsidian + holographic layer / leCore). Demo: https://brain.stacc.bio

Live

lecore-flash

Private

A founder who can live in HTML in the local memory and retrieval lane, taking a proved body rather than a blank repo.

Live

supercontext

Reproducible benchmarks for leCore retrieval: NIAH past the context window, MTEB, and cost-to-first-answer vs FAISS