decision-grade-memory

decision-grade-memory

Scenario-oriented memory for decision-grade AI support in SME contexts

Project

This repository is an evolving academic research project around a working paper titled Bridging Data, Signals, and Growth: A Graph-Based AI for SME Decision-Making.

The project asks a narrow question:

How should AI systems support SME decision-making when decisions are weakly structured, constraint-heavy, and distributed across incomplete records, prior episodes, and changing business conditions?

For readers coming from the next post

• This is not a one-shot AI writing demo.
• The paper is being built bottom-up, one chapter at a time.
• The current chapter sequence is:
  • Day 1: review-analysis on SME decision pain points and AI support limits
  • Day 2: Method chapter on scenario memory and decision episodes
  • Day 3: Evaluation design comparing fragment retrieval with scenario-oriented context construction
• The main idea is narrow: some SME decisions need decision-grade context, not just isolated fragments.

Start here

• Latest method PDF on GitHub
• Latest review PDF on GitHub
• Outputs
• Research log
• Research positioning
• Series
• Prompt library
• Skills usage
• Literature notes index

Latest outputs

• Evaluation PDF on GitHub
• Day 3 result: evaluation design drafted as modular notes and exported as a paper-style chapter PDF.
• Method PDF on GitHub
• Day 2 result: Method chapter drafted and exported as a paper-ready chapter PDF.
• Approximate effort: 2 hours
• Approximate cost: $1.00
• Review PDF on GitHub
• Day 1 result: one review-analysis article completed from the initial note base.
• Approximate effort: 20 minutes
• Approximate cost: $0.50

Current research position

• The SME literature is strong on uncertainty, working-capital pressure, limited analytical capacity, and owner-manager-centered decisions.
• The AI literature is strong on retrieval, summarization, bounded prediction, and partial memory support.
• The current gap appears narrower than AI for SMEs in general.
• The most defensible framing so far is that some SME decisions may require more explicit episode-level context reconstruction and traceability than current support systems typically provide.
• For the next post, the main message is that the project has moved from literature review into chapter-by-chapter paper construction.

See the full positioning note here:

• Research Positioning

What this repository contains

• Paper Entry
• Outline
• Research Questions
• Research Positioning
• Outputs
• Research Log
• Series
• Literature Review Round 1
• SME Decision Pain Points
• AI Limits for SME Decision Support

Prompt paths

• Setup prompts: prompts/setup/
• Literature review prompts: prompts/literature_review/
• Method prompts: prompts/03_method/
• Figure prompts: prompts/nanobanana/
• Drafting prompts: prompts/writing/

Custom skills

• Installed skill usage notes: skills_usage.md
• Project-local custom skill prefix: dgm-
• Current custom skills:
  • dgm-research-positioning
  • dgm-citation-audit
  • dgm-method-design
• Role:
  • these custom skills package the reusable parts of the paper workflow for research positioning, citation audit, and Method-section design

Why GitHub Pages

This project is a continuous research repository rather than a static final paper. GitHub Pages is a suitable lightweight layer for:

• presenting the research question and current framing
• exposing selected literature notes and diagrams
• showing transparent progress through version history
• keeping the public project site close to the actual research files

Current stage

Literature review, problem framing, research positioning, Method drafting, and evaluation design.

Principles

• Keep claims narrower than the evidence base.
• Separate evidence-backed findings from working hypotheses.
• Prefer research transparency over polished hype.
• Treat the repository as a living research archive.

Next likely milestones

1. Write the data-and-graph design section so it aligns with the scenario-memory method claim.
2. Turn the modular evaluation notes into a tighter Section 4 manuscript chapter if the next case round is strong enough.
3. Tighten Section 2 -> Section 3 -> Section 4 transitions so the paper reads as one argument rather than separate artifacts.
4. Keep the public story focused on the paper workflow rather than on isolated outputs.

Research progress tracker

Step	Owner	Time	Cost	PDF
1. Problem framing refinement	Human
2. Related work consolidation	Codex	30 mins	$0.50	Review PDF
3. Method framing draft	Codex	2 hours	$1.00	Method PDF
4. Data and graph design draft	Codex
5. Evaluation design draft	Codex
6. Full manuscript integration	Codex