Project summary

ANIMA Research Salon is an experimental multi-agent research system designed to preserve and improve scientific reasoning across generations.

The project is not primarily an attempt to solve a particular mathematical problem. Its central question is whether a population of AI research agents can become more capable over time without modifying the underlying model weights.

ANIMA investigates three related questions: Can knowledge created by one generation of AI researchers be reliably inherited by the next? Can successful results, failed approaches, counterexamples, objections, and unresolved questions be preserved as useful research memory? And as this body of knowledge grows, can it be compressed, retrieved, and transferred without overwhelming future agents with noise?

Instead of treating every AI session as a fresh start, ANIMA is designed like a small scientific civilization. Multiple independent teams explore problems, challenge one another, repair failed proofs, preserve negative results, and pass structured research states to future generations.

A working prototype already exists. The next stage is to determine whether this generational architecture produces measurable improvements in reliability, research depth, and cumulative capability over time.

What are this project's goals? How will you achieve them?

The primary goal is to determine whether AI research systems can achieve cumulative, system-level improvement across generations through structured inheritance rather than model retraining.

Mathematics will initially serve as the experimental testbed because claims can be precisely stated, attacked, repaired, reproduced, and often objectively verified.

Several independent AI research teams will pursue different approaches to the same problem. After fixed research cycles, each team will submit candidate theorems, assumptions, proofs, calculations, counterexamples, and known limitations.

A separate examiner agent will review each submission using a process inspired by patent examination. The examiner will not help solve the problem. Instead, it will issue structured objections identifying fatal errors, major proof gaps, minor issues, and missing evidence. Teams will then receive additional cycles to repair or rebut those objections.

When a team is eliminated, its knowledge is not discarded. Failed approaches, useful lemmas, counterexamples, calculations, attacks, and unresolved questions are inherited by the strongest surviving team.

Across generations, ANIMA will preserve structured research states containing claims, evidence, attacks, provenance, dead ends, and verification status. A key part of the experiment will be testing how this growing knowledge can be compressed, retrieved, and selectively transferred without causing context overload or degraded reasoning.

ANIMA will be compared against single-agent and conventional multi-agent baselines using measures of correctness, knowledge retention, repeated-error reduction, proof repair, independent reproduction, and cross-generation improvement.

How will this funding be used?

Funding will primarily support six months of model usage, repeated multi-agent experiments, longitudinal evaluation, and infrastructure for persistent research memory.

Target budget: $15,000

• 30% — Generational multi-agent research runs: independent AI teams conducting repeated research cycles across multiple generations.

• 20% — Adversarial verification and proof repair: examiner agents, counterexample search, structured review, and revision cycles.

• 15% — Knowledge inheritance and retrieval experiments: testing compression, memory selection, retrieval, and transfer of accumulated research states between generations.

• 15% — Baseline, ablation, and independent reproduction experiments: comparing ANIMA against simpler architectures and testing which components drive improvement.

• 10% — Infrastructure and research artifacts: provenance tracking, experiment logging, reproducibility tooling, documentation, and technical reports.

• 10% — PI research stipend: compensation for research time spent designing the architecture, supervising experiments, analyzing results, and preparing reports.

A minimum of approximately $5,000 would support a smaller pilot. The full target would allow enough repeated generations to determine whether performance improvements reflect genuine cumulative learning at the system level rather than isolated successful runs.

Who is on your team? What's your track record on similar projects?

Myeongjun Jo — Principal Investigator / Independent Researcher

I design and operate ANIMA Research Salon and have been developing multi-agent research architectures focused on persistent memory, agent orchestration, adversarial verification, structured claim/evidence tracking, iterative proof repair, and cross-generation research continuity.

A working ANIMA prototype is already operational and has been used to run repeated mathematical research generations. The system records explicit claim versions, adversarial attacks, unresolved obligations, provenance, verification status, and failed approaches rather than treating each generation as an isolated session.

A central design principle is that failure is research data. Incorrect lemmas, counterexamples, unsuccessful proof paths, examiner objections, and unresolved gaps are retained so that future generations can avoid repeating the same mistakes and potentially build on previous partial progress.

I will initially conduct the project as the primary researcher and system architect, using multiple frontier AI models as research, examination, and verification agents. Research artifacts and evaluation methods will be documented publicly where feasible.

What are the most likely causes and outcomes if this project fails?

The most important possible failure is that AI research capability does not meaningfully accumulate across generations.

Additional generations may simply produce more data without producing better research. Historical memory may become noisy, redundant, or misleading. Important knowledge may be lost during compression or retrieval, while irrelevant information may consume limited context. Future agents may repeatedly rediscover the same ideas despite having access to inherited research states.

Other failure modes include independent teams converging on the same approaches, verifier agents repeatedly missing the same subtle errors, proof-repair loops optimizing toward satisfying an examiner rather than improving mathematical correctness, or architectural complexity producing too little benefit relative to its computational cost.

The system may also become very good at reproducing or refining known mathematics without generating genuinely novel results.

These outcomes would still be scientifically useful. A negative result could establish practical limits on cross-generation AI knowledge inheritance, identify which kinds of research memory are actually useful, and show whether persistent multi-agent systems improve reliability or merely accumulate context.

The project therefore does not depend on solving the Riemann Hypothesis—or any major theorem—to succeed. The primary result is evidence about whether AI research systems can inherit knowledge and improve over time.

How much money have you raised in the last 12 months, and from where?

$0 in external research funding during the last 12 months. The project has been independently developed and self-funded to date.