Read cofounder Muhammad Umar's memo, Compression and Reinflation
CodeNinja AtomsUnited StatesSaudi ArabiaPakistan

Why it compounds

Context is the moat.

Every operation can host the same open weights models. Context is what compounds.

01 CAPTURE

Engram

Dedicated inference for your sovereign AI models.

ProductHyper EngramTrainsSovereign AI modelsIntoWeights and memoryRunsOn premises · air gapped
§00Overview

Train sovereign AI models in their weights and in their memory, on premises and in air gapped environments.

Hyper Engram turns the work an operation does into training signal for the open weights models it hosts. Runs, outcomes and corrections are captured as episodes, compressed into versioned principles and applied two ways: post training into the model's weights, and agentic memory the model reads at inference. Training data, weights and memory never leave the perimeter.

§01The learning loop

Recursive improvement, one loop at a time.

A master agent runs each RSI Loop. Two routines distil and verify what was learned every day. They never call each other directly: each reads and writes one shared memory, which the model reasons over at inference.

gap records, asks, demandnewest episode, open asksversioned briefsfeedback, open asksverified entriesmemory, mirroredplans, unmet asksmeasuresmissescode, tests, deploysepisodesfeedbackgap recordsbriefscorpusENGRAMMaster agentruns each RSI Loop, changes the codeDistillation routineevery morning: learnings into briefsCorpus routineevery morning: measures, lands entriesPlannerreads memory the moment it reasonsBenchmark and premortemfresh tasks and blind sets
Sheet 01Recursive learning loopHyper Engram
AReadWhere the last loop left the platform and both routines.
BPlanA short ranked list of changes, each with its measure.
CBuildSmall commits, every release tagged and reversible.
DReal requirementsDesigns on documents nobody wrote for a test.
EBenchmarkFresh tasks against frontier AI on its own.
FPremortemWould the design have caught a real incident?
GFeed the routinesEvery finding written into memory for its owner.
HCloseWhat the loop learned, read first by the next one.
§02Platform

Four stages. One training platform.

Capture, compress, train and reinflate. Every stage runs inside your perimeter, on premises or air gapped, on models you host.

01

Capture

Every run becomes training data. Each run is kept as an episode with its outcome, its feedback and the corrections a reviewer made, inside your perimeter.

02

Compress

Distillation into principles. Daily routines distil raw learnings into versioned principles that cite their sources. Contradictions are resolved and what never mattered fades.

03

Train

Into the weights. Curated episodes and principles become post training data for the open weights models you host, trained on your own hardware.

04

Reinflate

Into memory, at inference. The model reads the compressed memory and expands it into the problem in front of it: this site, these regulations, these people.

§03Reinterpretation

Learned tonight. Used tomorrow morning.

Each routine opens its run by reading the feedback and gap records on its newest episodes. A rule recorded today changes tomorrow's behaviour, with no code change.

  1. A run misses something

    A design senses the warning a real incident would have started with, and never says who must act on it.

  2. A later reader finds it

    The master agent reads the run and files a gap record on its episode: what was missed, the fix, and a rule for next time, written in the routine's own terms.

  3. The routine reads before it acts

    Its intake lists every rule in force. The rule is now part of how it works, with no change to its code.

  4. The record closes, or it does not

    The next RSI Loop checks whether the gap closed. A record still open is the first item of the next plan.

§03One gap record
// filed on the routine's newest episode
{
  gap: "designs sense the warning but never name who acts",
  fix: "brief a routing principle from general practice",
  found_by: "master agent",
  status: "open",
  rule_for_next_time: "every sensed warning names the role
    that acts, the reading that starts the clock
    and the options of the decision"
}
EPISODEWhat the run did and measured.
FEEDBACKWhat was wrong with how it worked, with a verdict.
GAP RECORDSWhat a later reader found it missed.

Every learner in the system, from the routines to the fleet, keeps its episodes in this one shape, so any reader can correct any writer.

§04Compression bench

Compress once. Reinflate anywhere.

Eight raw learnings are compressed into one principle, which the model reinflates into designs for operations it has never seen.

Experience8 learnings · 4 sources
  • Benchmark missA conveyor design sensed belt misalignment and never said who acts on it.
  • Engineer correctionThe alarm went to the control room, which cannot stop a haul truck. Route it to the pit supervisor.
  • Real requirementA tender asks for alerts to reach “the responsible officer” within a minute. Responsible for what?
  • Field learningOperators ignore an alert that arrives without a decision attached to it.
  • Benchmark missA gas reading crossed its limit. The design logged it and paged no one.
  • Engineer correctionName the reading that starts the clock, not only the threshold.
  • Earlier noteSend every alert to the shift manager.
  • Earlier noteMake critical alerts red.
Briefrouting · v6

Press Compress. Six learnings become one principle, one contradiction is resolved and one note fades because it never mattered.

Reinflate intothe model reads it here
Designno place yet
Waiting
Compress, then pick a place. The same principle becomes a different design in each one, because the world it lands in is different.

Illustrative. The principle is one the platform learned this way; the learnings and places are drawn for this page.

Sheet 04Compression and reinflationIllustrative
§05Weights and memory

Two training targets. One platform.

Engram places each lesson where it does the most good. Stable skills used at volume are trained into the weights. Knowledge that changes with the operation is kept in memory, where it stays current, readable and reversible.

Into the weightsInto memory
Best forStable skills used at high volumeKnowledge that changes every month
A new regulation arrivesJoins the next training runOne new entry, verified and live within hours
Reading what was learnedMeasured through evaluationEvery principle lists the learnings it came from
A stronger base model arrivesThe training run is repeated on itThe memory carries over on day one
Undoing a lessonRoll back to an earlier checkpointRetire one versioned principle
Where it runsInside your perimeter, on models you hostInside your perimeter, on your infrastructure
§06Evidence

Measured every loop.

Praxis, trained on Engram, is measured on PADI, our benchmark for physical AI system design, against frontier AI on its own, on fresh tasks every RSI Loop.

48.5%

Praxis, RSI Loop 22, blind

28.4%

Frontier AI on its own, same tasks

7of 7

Unseen tasks where Praxis came out ahead

13

Industries measured, in three countries

Praxis platformFrontier AI on its own
0%25%50%75%100%1213141516171819202122 blind48.5%28.4%RSI LOOP

Checks passed as a share of checks judged, each RSI Loop. Loop 22 is a blind evaluation on tasks the platform had never seen; earlier loops reused some tasks and show the trend. Results are public; the method stays private. Every figure on PADI →

§07Principles

Eight rules every loop keeps.

They hold across every phase and every agent. A change that would bend one stops for human review.

01

Train where it belongs

Stable skills go into the weights. Knowledge that changes goes into memory.

02

Meaning, never keywords

Any decision about what a sentence means is put to a model and stored. Keyword scoring is banned.

03

Every claim stands on a source

An entry needs a dated source. A principle needs the learnings it was distilled from.

04

Evaluation stays blind

Held out tasks are never shown to the planner. Incident cases never feed the memory.

05

Two agents, two views

Every writer has an adversarial verifier. Nothing verifies its own output.

06

The record is the context

Commits, episodes and gap records carry the why. A session that lost its memory reads the log first.

07

Every release is reversible

Each RSI Loop is tagged. A bad change is reverted by its own commit, a bad principle by its own version.

08

Sovereign by design

Training data, weights and memory stay in your perimeter, on premises or air gapped.

ProjectHyper Engram
TrainsWeights and memory
Talk to us →Read the memo →

What is Hyper Engram?

Hyper Engram is CodeNinja's platform for training sovereign AI models on premises and in air gapped environments. It trains open weights models in their weights and in their memory, from the runs, outcomes and corrections of the operation they serve.

How does Hyper Engram train a model?

Runs are captured as episodes and compressed daily into versioned principles that cite their sources. Curated episodes and principles become post training data for the model's weights, and the same principles form an agentic memory the model reads at inference.

What is compression and reinflation?

Compression turns many observations into one small principle, resolving contradictions and dropping what never mattered. Reinflation is the model expanding that principle into a new situation: a different country, operation or set of people. The idea follows Michael Levin's work on self improvising memory in living systems.

What goes into the weights and what into memory?

Stable skills used at high volume are trained into the weights, where they are fastest and cheapest to serve. Knowledge that changes with the operation, such as regulations, equipment and procedures, is kept in memory, where it is current within hours and can be read and reversed.

What happens when a better base model is released?

It is swapped in. The memory carries over unchanged, and the curated training data is applied to the new base model.

Where does Hyper Engram run?

Inside the customer's perimeter: on premises, in a private cloud or fully air gapped, on models the customer hosts. Training data, weights and memory never leave it.