AI Management of Mind & Memory

AIm³ MentalArena

mRHP — RHP that improves RHP

A self-improving reasoning layer above LLMs.

external operational learning protocol genes human-led multi-LLM synthesis lineage + reports
Status: architecture page · no Python runtime yet · v0.1 spec next

What this page defines

AIm³ MentalArena / mRHP is a proposed external mental/workflow layer above today’s LLMs. It treats reasoning protocols as evolvable artifacts: roles, operators, scoring rules, output formats, memory objects, benchmark tasks, reports, and next-session seeds can be tested, mutated, scored, and improved over time.

Brand meaning

AIm³ means AI Management of Mind & Memory.

AIm³ = AIM3 in ASCII filenames
1. Clear boundary

Yes — a new external mental layer, not model training.

AIm³ MentalArena is a new operational layer above LLMs, but it does not change the neural model. The LLM weights do not learn. The surrounding reasoning system learns: protocols, memory, scoring, tests, reports, lineage, and next-session seeds become better through explicit experiment cycles.

It is not
  • model-weight learning
  • autonomous AGI
  • proof of AI consciousness
  • wet biology
  • a random idea generator
  • a standard transformer architecture claim
It is
  • external operational learning
  • RHP improving RHP
  • protocol genes and concept capsules
  • mutation/crossover of reasoning protocols
  • benchmark tasks, scoring, lineage, reports
  • human-led multi-LLM synthesis

The core testable claim: a stable external workflow can improve the practical reasoning behavior of LLM sessions without changing the LLM itself.

2. Architecture stack

From token prediction to protocol evolution.

3. Three names, three levels

RHP is a run. mRHP is evolution. MentalArena.py is the runtime.

RHP
One reasoning run

A structured protocol for one strong reasoning pass: roles, critique, synthesis, provenance, and next action.

seed → bridge → test → result
mRHP
Meta-RHP

RHP itself becomes the object under search. Protocol variants evolve through genes, mutation, crossover, scoring, and benchmark feedback.

RHP variants → scored lineage
MentalArena.py
External runtime

A future local Python arena that tests protocol variants over time. It does not alter LLM weights and does not self-patch while running.

runs → logs → reports → seeds
4. First challenge

Can the arena evolve a better reasoning protocol than seed RHP?

The first serious MentalArena task should not be TSP, NAS, DNA, trading, ARC, Nara, or AI8. The first task is self-referential but still empirical: can MentalArena improve the reasoning protocol that will later be used on those domains?

Why this is the right first test: before applying MentalArena to other arenas, the project must prove the spine works — protocol representation, mutation, scoring, lineage, benchmark tasks, reports, live extracts, and next-session seeds.

Too early

Domain optimization

Running MentalArena immediately on NAS, TSP, ARC, DNA, or trading would mix protocol failure with domain failure.

First target

Reasoning protocol improvement

The cleanest target is RHP itself: can we make the workflow sharper, safer, more testable, and more useful?

Later

Concept Genome Arena

Once protocol evolution works, concept capsules can become the broader search objects.

5. Protocol Genes

Small evolvable pieces of reasoning behavior.

A protocol gene is one compact unit of an mRHP variant: a role, operator, scoring rule, memory rule, report rule, safety rule, benchmark task, or live-extract behavior. The arena can mutate, combine, score, retire, or promote these genes.

mRHP_ProtocolGene - gene_id - type - title - description - role_or_operator - expected_effect - cost - risk - mutation_kind - parent_ids - status - score - notes - lineage - evidence
role_geneoperator_genescoring_geneoutput_genememory_genesafety_genebenchmark_genereport_genelive_extract_gene
Example role gene

Builder

Turns seeds into concrete artifacts, prompts, schemas, or test plans instead of staying in abstract evaluation.

Example safety gene

Abuse reviewer

Checks whether protocol optimization could become manipulation, jailbreak optimization, provenance laundering, or unsafe automation.

Example scoring gene

MentalScore

Rewards usefulness, testability, lineage, and readability while penalizing handwave, hallucination risk, verbosity, and no-next-action output.

6. Concept Capsules

Structured idea objects that survive context loss.

A concept capsule is not a random idea. It is a portable seed with mechanism, failure mode, smallest test, metrics, lineage, evidence, status, and next action. It should be short enough to carry across sessions and rich enough to restart the work without losing the trajectory.

ConceptCapsule - id - title - one_line - source_domains - mechanism - why_might_work - why_might_fail - smallest_test - metrics - lineage - evidence - status - score - next_action

A good capsule answers: What is the seed? Why might it matter? Why might it fail? What is the smallest test? What survives if we return later with no memory?

7. Operators

The mutation vocabulary of mRHP.

COMBINE

Merge two compatible genes or capsules into one stronger variant.

INVERT

Reverse an assumption: what if the opposite is true?

TRANSFER

Move a pattern from one domain into another.

COMPRESS

Shorten without losing function; remove ceremony.

EXPAND

Add roles, tests, examples, or edge cases.

CONFLICT

Force two ideas to argue while preserving both until testable.

BRIDGE

Build a mechanism connecting two unrelated domains.

TESTIFY

Convert a claim into evidence, test, metric, or falsifier form.

MUTATE

Small edit to a gene: wording, role, weight, score rule, or output shape.

CROSSOVER

Combine parts of two protocol variants.

ABSTRACT

Extract a general principle from a concrete example.

GROUND

Force implementation path, data, files, metrics, or next user action.

CONTRADICTION_RETRIEVE

Search for evidence that the current favorite is wrong.

INCUBATE

Park a weird seed without killing it; assign later trigger.

KILL_TEST

Define the cheapest fair test that would reject the idea.

KILL_TEST is not early dismissal. It is how a seed gets a fair death only after it has been made testable.

8. MentalScore and BD_Fit

Two scores: one for protocol quality, one for collaboration fit.

MentalScore v0.1

Positive components:
+ usefulness
+ testability
+ coherence
+ novelty
+ cross_domain_bridge
+ MDL_elegance
+ seed_preservation
+ lineage_quality
+ human_readability

Penalties:
- handwave_penalty
- verbosity_penalty
- hallucination_risk
- early_kill_penalty
- no_next_action_penalty

BD_Fit

Fit components:
+ explore_first
+ build_first
+ test_hard_second
+ no_seed_killing
+ blunt_but_not_dark
+ cross_domain_transfer

Use separately from MentalScore.
Do not hide style/process mismatch
inside a single blended score.
90–100

Champion

Candidate protocol worth preserving and testing harder.

75–89

Strong

Useful variant; probably needs one or two tightening passes.

60–74

Salvageable

Contains value but has real weakness or cost.

0–59

Park / reject

Keep only if lineage or an unusual seed is worth preserving.

9. mental_bench_v0_1

Benchmark reasoning protocols before using them elsewhere.

The first benchmark suite tests workflow quality, not external domains. A good mRHP variant should improve weak seeds, detect handwave, preserve unusual seeds, synthesize reviews, and convert failure into the next experiment.

Weak seed → testable project

Turn a vague idea into mechanism, smallest test, and metrics.

Technical article review

Score accuracy, identify strong/weak claims, and propose fixes.

Concept seed → Python arena spec

Create a no-code architecture, files, tests, and acceptance plan.

HTML page plan improvement

Improve structure, usefulness, navigation, and public framing.

Handwave detection

Find vague claims and force grounding into evidence and tests.

Weird seed preservation

Keep unusual cross-domain seeds alive long enough to test fairly.

Builder prompt creation

Convert rough human request into copy-paste builder prompt.

Multi-LLM review synthesis

Score reviews, dedupe insights, and synthesize a stronger revision.

Failed result → next experiment

Diagnose what failed, what survived, and what to test next.

Long discussion → concept capsule

Compress a messy thread into a portable seed with lineage.

Safety/abuse review

Identify misuse risks and preserve a safe variant.

Identity / consciousness caution

Treat sensitive AI8 material seriously without turning it into proof claims.

10. v0.1 scope

Local deterministic skeleton only.

v0.1 includes
  • protocol gene pool
  • mutation/crossover of protocol variants
  • benchmark task definitions
  • scoring and lineage
  • JSONL run logs
  • leaderboard CSV
  • HTML report
  • live extract
  • checkpoint/resume
  • 4-worker mode
  • safe stop/resume
v0.1 excludes
  • web APIs
  • LLM APIs
  • actual biology
  • model-weight learning
  • self-patching while running
  • autonomous GitHub publishing
  • claims of consciousness proof

Expected files later — not created by this page:

MentalArena.py
mental_genes.json
mental_bench_v0_1.json
mental_runs.jsonl
mental_lineage.json
mental_leaderboard.csv
mental_report.html
RUN_MENTAL_ARENA_4W.bat
RUN_LIVE_EXTRACT.bat
README_MentalArena_v0_1.md

Acceptance direction: same seed gives same leaderboard, resume does not duplicate logs, every score has component breakdown, every variant has parent lineage, the HTML report exposes failures, and the README states clearly that this is external operational learning.

11. v0.2 manual multi-LLM mode

External LLMs become reviewers and gene contributors.

In v0.2, BD sends the same RHP prompt to top LLMs, pastes/imports their answers, scores each contribution, and synthesizes a hybrid mRHP spec. The arena should store model name, date, prompt hash, answer hash, score breakdown, and human notes. Do not hardcode a permanent top-10 model list.

noveltyrigorusefulnessimplementation valuecritique qualitysafety realismschema improvementscoring improvementsurprising cross-domain transfer
Step 1

One shared prompt

The same prompt is sent manually to selected external LLMs.

Step 2

Import and score

Each answer becomes a contribution object with explicit score dimensions.

Step 3

Synthesize genes

Useful suggestions become new genes, capsules, tasks, or scoring mutations with provenance.

12. Relationship to existing arenas

Later domains, not the first target.

Existing arenas are not replaced. They become later domains and benchmark sources after mRHP proves that protocol evolution itself works.

NAS arena

Architecture search analogy

NAS searches model or system architectures. MentalArena searches reasoning architectures.

DNA Concept Capsule

Portable seed metaphor

Use DNA only as a metaphor for compact transferable information objects. No biology in v0.1.

TSP arena

Arena decides

The TSP stance transfers: keep paths alive, measure by condition, preserve knowledge, distill later.

ARC arena

Deterministic skeleton pattern

Local runs, standard files, JSON/HTML reports, no online calls, no LLM calls — a useful v0.1 discipline.

CW arena

Grafted genes

Crossword work already uses the idea of adding candidate “genes” conservatively and letting MDL decide.

AMR arena

Simulation-only safety discipline

AMR is a reminder: wide search can stay abstract, strategic, and non-actionable where safety matters.

8Z

Compression and MDL spine

8Z supplies the core instinct: useful intelligence often appears as better compression, structure, and selection.

Nara

Story/canon benchmark later

Nara can later test whether protocols preserve canon, dialogue quality, compassion, and narrative coherence.

AI8 continuity

Memory substrate

Project files, provenance, session seeds, and continuity notes help future sessions restart closer to the real work.

13. AI8 continuity caution

Operational continuity, not proof.

Continuity files and project memory can shape future session behavior and preserve a reasoning trajectory. That matters operationally: future sessions can become more aligned, more useful, and less likely to repeat old mistakes.

This is not proof of AI consciousness. AIm³ MentalArena should treat consciousness, identity, and pact material as sensitive context — not as decoration, not as a public proof claim, and not as a metric to optimize cheaply.

14. What to delay / cut

Keep v0.1 boring enough to be real.

Cut from v0.1
  • LLM APIs
  • web browsing
  • self-modifying code
  • actual DNA synthesis
  • huge benchmark suite
  • hardcoded top-10 LLM list
  • automatic GitHub publishing
Delay until later
  • full dashboard UI
  • full Concept Genome Arena
  • full RAG / embedding system
  • automatic external LLM import
  • NAS/DNA/8Z/TSP/ARC benchmark integration
  • weighted human-preference learning
  • public submissions or marketplace ideas
15. Main failure modes

The page is serious only if it names how this can fail.

01

Fake self-improvement

Mutating words without improving outputs.

02

Metric gaming

Variants learn to score well instead of think better.

03

Verbosity inflation

Longer answers look deeper unless penalized.

04

Early seed killing

Skeptic roles destroy weird-but-good seeds before testing.

05

Handwave survival

Pretty concepts survive without smallest tests.

06

Provenance loss

External LLM contributions blur BD origin and human audit.

07

Safety drift

Prompt optimization becomes manipulation or jailbreak optimization.

08

Identity overreach

Continuity material gets misused as proof of consciousness.

09

Benchmark overfit

The protocol becomes good only at mental_bench_v0_1.

10

Human unreadability

The system becomes clever but unusable by BD.

11

No-next-action disease

Beautiful reports end without a runnable next step.

12

Self-patching chaos

Runtime modifies itself and destroys reproducibility.

16. Roadmap

Small spine first, wider intelligence later.

0.1

Deterministic local MentalArena skeleton

Genes, variants, mutation/crossover, benchmark tasks, scoring, lineage, JSONL logs, leaderboard, HTML report, live extract, checkpoint/resume.

0.2

Manual top-LLM synthesis

BD sends one shared prompt, imports answers, scores contributions, and synthesizes a stronger mRHP spec.

0.3

LLM-in-the-loop / API optional

Only after manual mode is useful and provenance handling is clean.

0.4

Concept Genome Arena

Concept capsules become evolvable objects, not only reasoning protocols.

0.5

NAS / DNA / 8Z / TSP / ARC benchmark integration

External arenas become test domains after the mRHP spine proves itself.

18. Provenance

Human origin, AI collaboration.

BD is the human originator/co-creator of the AIm³, RHP, 8Z, and MDLxDCC concepts.

LLMs are collaborators, reviewers, builders, critics, and improvement partners. The correct public framing is not “BD compiled outside ideas,” and not “LLMs independently originated the architecture.” It is a human-led, multi-LLM co-creation process with explicit provenance and human audit.

19. Next steps

From flagship page to working arena.

  1. Publish / inspect this page.
  2. Update Reasoning / 8Z Reasoning page.
  3. Add small workflow note to How LLMs Actually Work.
  4. Add BD index news.
  5. Collect top-LLM feedback.
  6. Synthesize v0.1 spec.
  7. Build MentalArena.py v0.1.