Enterprise adoption of AI agents will soon face a fundamental obstacle: LLMs cannot persistently retain company-specific knowledge, experience, and decision histories in a verifiable form. Larger context windows do not solve this problem. In the post-AI-bubble era, the central development challenge will shift toward building Operational World Models.
Mindware Multi-Client Development Project
Phase 1 Development Plan (Draft)
Building on LLM Wiki and Concept Index
Building a Self-Organizing Operational World Model
through Organizational Knowledge Mining
Phase 1: October 1, 2026 – September 30, 2027
Organizer: Mindware Research Institute, Inc.
The system will ingest information scattered across internal documents, relational databases, operational logs, reports, and other sources; extract candidate elements with LLMs; support human validation; form concepts through ConceptMiner and Concept Index; and structure the results as Episode Cards and process models. The outcome will be an enterprise-governed Operational World Model shared by human users and AI agents.
- Intended Participants
- Executives and practitioners responsible for AI/DX, information systems, R&D, corporate planning, new business, and knowledge management across manufacturing, distribution, retail, telecommunications, IT, finance, food and beverage, and other industries
- Participation Fee
- JPY 2.5 million per unit, excluding tax / up to two designated participants per unit
- Payment Terms
- 50% upon application and contract execution; 50% before the project begins
- Participation Format
- Multiple units permitted; company and participant names confidential in principle
- Launch Condition
- No minimum number of companies; the project will proceed from one participant
- Initial Expansion Target
- 20 companies; additional development, external specialists, and evaluation environments will expand with participation
Project Overview
This is a one-year, multi-client development program in which participating companies jointly support the creation of a common software foundation that transforms documents, data, logs, conversation records, and decision histories scattered across an organization into reusable knowledge, experience, and process assets. The primary objective is not company-specific research or consulting, but the establishment of a reusable product foundation applicable across multiple organizations.
Company-specific knowledge does not persist
Chat histories and summaries do not become governed enterprise assets with source traceability, change history, access control, and outcome evaluation.
RAG alone cannot model operational reality
Document retrieval cannot by itself represent who made what decision, under which circumstances, with what outcome, or how work proceeds through sequences and branches.
Analytical outputs are not operational artifacts
Clusters and process maps are not enough. The system must produce manuals, case collections, decision rationales, and improvement candidates that business users can continue to use.
Development Objective: Building an Operational World Model through Organizational Knowledge Mining
The Operational World Model proposed here is not a massive model intended to simulate the world in general. It is an organization-specific model that continuously updates actors, objects, states, events, decisions, outcomes, relationships, and operating procedures with traceable sources.
World ModelModel how the organization actually operates from dispersed operational information
Core Development Outcomes of Phase 1
Phase 1 will connect the following five outcomes as one information-flow and governance foundation rather than as isolated experiments.
Organizational Knowledge Mining Core
Extract candidate actors, actions, objects, states, decisions, and outcomes from documents, free text, operational data, and logs, then convert validated candidates into reusable knowledge units.
Episode Memory Core
Accumulate situations, events, decisions, outcomes, evaluations, and sources as Episode Cards and retrieve related experience for reuse.
Process Mining Core
Transform structured logs and human-validated event candidates from unstructured information into models of actual workflows, exceptions, rework, and branching paths.
Operational World Model Layer
Provide a common schema and reference APIs linking document knowledge, concept nodes, episodes, processes, sources, and permissions for continuous updating.
Business Artifact Generator
Convert analytical outputs into drafts of operating manuals, process descriptions, case collections, decision rationales, and improvement candidates.
Core Workflow Enabled by the Completed System
Ingest documents, data, logs, and events from business systems, relational databases, files, and automation tools such as Zapier.
Preserve source information such as original text, record IDs, timestamps, authors, and system names for later verification.
Use LLMs as language-processing components to extract and normalize candidate actors, actions, objects, states, decisions, outcomes, reasons, and evaluations.
Allow human reviewers to validate candidates, correct errors, add evaluations, approve content, and define access scopes.
Generate knowledge pages, Episode Cards, event logs, and process descriptions from validated information.
Use ConceptMiner and Concept Index to form conceptual groupings, and process mining to identify actual flows, exceptions, rework, and branching paths.
Enable business users to work with the results as manuals, case collections, and process models, while AI agents retrieve relevant knowledge and experience through MCP or APIs.
Record new decisions and outcomes and continuously update the organization-specific knowledge, experience, and process models.
Minimum Scope of the Phase 1 Reference Implementation
The project does not promise fully autonomous agents or automated strategic decision-making. Priority will be given to completing an implementation foundation that participating companies can inspect and evaluate themselves.
Illustrative Use Cases
Quality and Maintenance
Information Ingested: Anomaly reports, equipment logs, causal hypotheses, responses, and recovery outcomes
Outputs Generated and Used: Similar-incident cards, effective and ineffective responses, deviations from standard procedures, and draft maintenance manuals
Research and Development
Information Ingested: Experimental conditions, design reviews, failures, decisions, outcomes, and next hypotheses
Outputs Generated and Used: Recall of failure cases, hypothesis lineage, conceptual clustering of conditions, and reconstruction of experimental records
Sales and Customer Service
Information Ingested: CRM records, sales-call notes, proposals, responses, and reasons for wins or losses
Outputs Generated and Used: Similar customers and opportunities, win-loss factors, response processes, and draft proposal and service playbooks
Manufacturing, Logistics, and Administrative Operations
Information Ingested: Work logs, application histories, emails, handoffs, and daily reports
Outputs Generated and Used: Actual workflows, rework, delays, duplication, process documentation, and improvement candidates
Management and New Business Development
Information Ingested: Meeting records, assumptions, options, decisions, execution outcomes, and evaluations
Outputs Generated and Used: Decision histories, consistency with past policies, outcomes of similar decisions, and collections of decision rationales
Skill and Knowledge Transfer
Information Ingested: Interviews with experienced personnel, scattered documents, cases, and tacit decision criteria
Outputs Generated and Used: LLM Wiki pages, business processes, case collections, records of exception handling, and training drafts
Why Enterprise AI Now Needs an Operational World Model
01
Large language models possess extensive general knowledge and language capabilities, but they do not preserve company-specific experience, decision criteria, or operational reality in a persistent and verifiable form.
02
RAG and vector search are useful for locating document fragments, but they cannot by themselves model decision histories or the sequences and branches through which work actually proceeds.
03
This project uses LLMs as standard components for extraction, summarization, verbalization, and dialogue. Its competitive focus is the mechanism that continuously turns enterprise observations into concepts, experience, and processes under human governance.
Avoiding the Mistakes of the KDD and Data-Mining Era
Even a useful discovery algorithm becomes difficult to use when the distance between its output and a practical business artifact is too great. This project includes the conversion from discovery into documents, processes, and cases within the product itself.
Existing Implementation Foundation
This project does not begin from zero. Since April 2025, Mindware Research Institute has developed and validated the following technologies as working systems.
ThinkNavi / LLM Wiki
ImplementedReconstructs documents as knowledge pages, routes questions to appropriate pages, and references source documents. Phase 1 will separate Single Wiki Builder and use it as the foundation for business-artifact generation.
ConceptMiner
ImplementedApplies Growing Neural Gas and a Minimum Spanning Tree to embedding space to form concept nodes and relationship structures.
Concept Index
In DevelopmentMaps records in existing relational databases to concept nodes, enabling conceptual extraction through SQL and connecting operational data and episodes to the model.
CCM / Thinking Support
ImplementedAn experimental environment for associative responses and decision support using relationships among concepts, episodes, and decision candidates.
Core Philosophy: A Self-Organizing Concept Approach
Grow concept structures from observed data instead of defining a fixed ontology first
Rules, objects, and ontologies are useful for system integration, regulatory compliance, and access control. However, predefined structures alone cannot cover new similarities, exceptions, ambiguous judgments, and organization-specific meanings. The project forms concept structures from actual data while people name and formalize only what must be fixed.
Deliverables and Services for Participating Companies
Each participating company will receive the contractual foundation package regardless of participation scale. Twenty companies is not a launch condition; it is an expansion target for additional development and external review.
Organizational Knowledge Mining Core
Reference implementation covering information ingestion, source governance, LLM-generated extraction candidates, human validation, and generation of knowledge units and event candidates.
Episode Memory Core
Reference implementation including Episode Cards, related-case retrieval, API and webhook input, MCP and API access, and basic management functions.
Process Mining Core
Reference implementation including event-log generation, basic process discovery, visualization of paths, repetitions, and exceptions, and sample evaluation scenarios.
Operational World Model Specification
Common schema, API specifications, and sample data linking document knowledge, concepts, episodes, processes, sources, and permissions.
Single Wiki Builder
Participant-only source code extracted from ThinkNavi, including access to a restricted GitHub repository, setup instructions, and sample data.
Concept Index
Executable software and basic documentation for evaluating conceptual search. Supported operating systems, databases, and terms of use will be specified at delivery.
ThinkNavi / ConceptMiner Annual Usage Rights
Usage and evaluation rights for designated functions during Phase 1. API charges, processing scale, and usage scope will be defined in the participation terms.
Year-Round Online Program
Monthly sessions, technical demonstrations, development reports, an archive of recordings and materials, common-question support, and a final report.
Expansion According to Participation
Phase 1 Development Schedule (Draft)
October-December 2026
Core Work: Participant interviews, common-use-case definition, information ingestion, and data-model design for sources, permissions, Episode Cards, and event candidates
Main Verifiable Outcomes: Registration and editing UI/API, extraction and validation flow, and initial demonstration
January-March 2027
Core Work: Related-case retrieval, event-log generation, basic process discovery, ConceptMiner and Concept Index integration, and MCP server/API access
Main Verifiable Outcomes: Knowledge and episode retrieval, process visualization, and AI-agent access demonstration
April-June 2027
Core Work: Single Wiki Builder integration; generation of operational documents, process descriptions, and case collections; participant evaluation scenarios
Main Verifiable Outcomes: Integrated demonstration of knowledge, concepts, episodes, and processes
July-September 2027
Core Work: Incorporation of evaluation results, refinement of permissions and change history, documentation, samples, final tagged release, and outcome report
Main Verifiable Outcomes: Reference implementation, technical materials, evaluation results, and final symposium
Year-Round Online Program
Participation Terms and Project Positioning
Up to two designated participants per unit
Multiple units permitted
Payment Terms
50% upon application and contract execution; the remaining 50% before the project begins.
No Minimum Launch Threshold
The project will not be canceled and fees will not be refunded solely because participation does not reach the expansion target. Contracted core outcomes will still be delivered.
Participant Confidentiality
Company names, logos, and participant names will remain confidential in principle. Prior consent will be obtained before disclosure.
Intellectual Property and Governance
Participation does not confer ownership interests in intellectual property, voting rights, exclusive rights, or authority over the product roadmap.
Permitted Use
Deliverables may be used for internal evaluation and research within the contractual scope. Redistribution and use in competing services will be prohibited.
Company-Specific Work
Company-specific proofs of concept, integrations, data preparation, customization, and production deployment support will require separate agreements.
Research Foundations and Primary Reference Candidates
The project does not promise to reproduce each theory faithfully or implement all of them during Phase 1. These are design references for translation into reference implementations that can be evaluated in enterprise environments.
LLM Wiki
Andrej Karpathy
Compile source material into reusable, cross-referenced knowledge pages rather than searching raw fragments directly.
Process Mining
Wil M. P. van der Aalst and others
Discover actual business processes from event logs and analyze deviations, bottlenecks, repetitions, and exceptions.
Generative Agents
Joon Sung Park and others
Store experience as natural-language memory, retrieve it by relevance, recency, and importance, and use it for reflection and planning.
Complementary Learning Systems
McClelland, McNaughton, and O’Reilly
Link rapid recording of new events with slower formation of general structures from accumulated experience.
Case-Based Reasoning
Aamodt and Plaza
Retrieve, reuse, revise, and retain past cases as a continuing problem-solving cycle.
KDD / Knowledge Discovery
Fayyad, Piatetsky-Shapiro, Smyth, and others
Critically extend the KDD process by converting analytical outputs into documents, processes, and cases usable by business personnel.
Project Director
Beginning in 1982, he was involved in planning and operating technical seminars on advanced technologies, including new media, at a publishing and seminar company. From 1985, he conducted business-opportunity research in advanced technology fields at a major consulting company. In 1986, he organized an AI Chip Seminar that included fuzzy inference chips. In 1997, he wrote a serialized article on Concept Research for a management magazine. In 1998, he began research focused on Kohonen’s Self-Organizing Map. From 2000, he served as the Japanese distributor for Viscovery SOMine and also handled HUGIN and XLSTAT. Following XLSTAT’s acquisition in 2023, he began developing proprietary systems and established the core ConceptMiner technology using Growing Neural Gas and a Minimum Spanning Tree.
Request the Phase 1 Information Package and Implementation Demo
Detailed schedules, contractual terms, advisors, and final usage conditions for the deliverables will be confirmed after formal approval. Inquiries regarding participation, briefings, and implementation demonstrations are welcome.