spearmint ai · spare your machine bills

Cognition,
compiled.

A new kind of data: a model's understanding of a document, written down so that any model, any size can use it. Read once by the strongest model you have. Answered forever by the smallest one you can run.

Ask a compiled recordCompile your own
four walls, one cause

Raw data makes every
model re-read the world.

Tokens are a model's input format, not its working format. Every reader converts the same document into understanding again, from scratch, every time. That is where the walls come from.

Context window

Records scale to 10M tokens of source; around 1M is the sweet spot for an agent's working set. The reader never sees more than six thousand tokens per question, whatever the record holds.

Memory wall

A compiled record is memory that does not forget and does not drift: what was understood is stored, versioned and reloaded, not re-derived.

Data wall

Documents, logs, histories and canons compile into state a 0.6B model consumes directly. Capability moves into the data; the model can be small.

Token bills

The strongest model reads a document once. After that, each question costs a fraction of a cent on hardware you own, and the answer is cited.

360° view · your activity

Context overwhelm.
Token overwhelm.
Seen in one place.

Every record carries its own activity: how much source it holds, how little the reader needs per question, how many tokens never get re-read, and what that is worth. Live from the records on this site.

Ask a recordCompile your own

Records scale to 10M tokens of source. Around 1M is the sweet spot for an agent's working set.

My Activity

CONTEXT OVERWHELM · largest record
1.19Msource tokens
in one record
outer ring: the source against a 10M-token record ceiling · inner: what the reader actually holds per question
raw: 1,189,076 tokreader: ≤ 6,000 tok
windows: reader 32,768 · typical frontier 200,000 · this record 36× the reader's window
TOKEN OVERWHELM · per question
1.19M re-read, raw
with a record: ~6,000 · never re-read: 99.5%
RECORDS ON THIS SITE
—
— source tokens read once
PRE-RESOLVED REASONING
—
answers worked out before anyone asked
TOKENS NOT RE-READ · 300 questions each
—
what a re-reading pipeline would have billed, in tokens
the .mii record

Not a summary.
Not an index.
The understanding itself.

  1. Briefing and reading rules. What this is, who the actors are, the conventions, the traps. Always in the reader's state.
  2. Normalised facts. Every figure, date, term and clause, located in the original.
  3. Pre-resolved reasoning. Thousands of questions a professional would ask, already worked out with their arithmetic and evidence.
  4. Negative space. What the document does not say, so absence is an answer, not a guess.
  5. Reader state. The small model's working state, and a larger model's understanding projected into it, ready to load.
briefing2,632 tokens
Fourteen consecutive 10-Ks, FY2012 to FY2025. Fiscal year equals calendar year. Amounts in $ millions except per share…
facts11,135 cells
FY2019/c006.59 · Note 21 · "an earnings charge of $8,259 million, net of insurance recoveries of $500 million, in 2019"
pre-resolved reasoning12,775 rows
Q: how did the 737 MAX concessions hit the statements? → revenue reduction $8,259M; $6.3B into program inventory; ~$4.0B abnormal costs expensed 2020–21 [c006.59, c006.60, c003.77]
negative space1,938 facts
No filing from FY2012 to FY2025 expresses going-concern doubt; FY2020 and FY2021 assert funding is probable.
reader state4,411 slots · transferred
kv.reader · the 0.6B's own prefix state, with a 7B's understanding of the briefing projected in.
one format, every consumer

Anything that reads
can read a record.

Personal agents

Your mail, notes, calendars and history compiled once into a memory that survives sessions, models and devices.

Agents and tools

An agent carries the compiled record of its codebase, its runbooks or its case file and acts on it without re-reading.

Chatbots and characters

A persona's canon, voice and world compiled into state: consistent across every conversation, on a small model.

Enterprise review

Contracts, filings, clinical records, matters: read once, questioned forever, cited to the page, inside your tenant.

measured, not promised

A 0.6B reader on a record.
A frontier model on the original.

82.5
vs 82.1 for the frontier model, on human-written contract questions
80.5%
of 420 unseen free questions answered as the frontier would; retrieval over raw text: 13.8%
1.19M
tokens in one record, fourteen years of filings, beyond any model's window
~1 s
per question on one consumer GPU, for a fraction of a cent
live · the same 0.6B reader, right now

Ask anything.

See what others have compiled. Every record below was read once; click one to question it with the full derivation and the original passages.

loading…
compile your own

One key.
One document.
Forever answered.

A compilation key compiles one of your documents end to end in the demo: upload, watch the stages, and ask it anything within about 25 minutes for a typical contract. For a live corpus, a fixed-scope pilot:

  • Up to 5,000 documents compiled inside your tenant; records and the on-prem reader delivered to hardware you own.
  • Your taxonomy and question space folded into the compiler.
  • Parity measured on questions your own people write, against the frontier model you already pay for.
  • Five weeks. One number you can check.