Multi-Agent Knowledge Framework

A file-based knowledge management system for AI agent workspaces. Gives multiple AI agents (Claude Code, Codex, Gemini, local models, etc.) a shared, structured memory that compounds across sessions.

What this is

A protocol and directory schema — not a tool or library. Any AI agent that can read files can follow it.

• Retrieval protocol: agents know exactly where to look before asking the user
• Writing protocol: discoveries are filed to the right place automatically
• Knowledge graph: cross-references between files tracked in GRAPH.md
• Lint rules: weekly health checks keep the knowledge base consistent
• Raw staging layer: unprocessed sources land in raw/ before being compiled

Directory structure

.
├── AGENTS.md              # Retrieval + writing protocol (read this first)
├── CLAUDE.md              # Claude Code project instructions
├── agent-infra.md         # Your bot/model/billing config (private, gitignored)
├── knowledge/
│   ├── INDEX.md           # Master index of all knowledge files
│   ├── LINT-RULES.md      # Weekly health check rules
│   ├── GRAPH.md           # Cross-reference matrix (auto-regenerated)
│   ├── raw/               # Unprocessed source materials
│   ├── research/          # Investigation notes and feasibility studies
│   ├── guides/            # How-to runbooks
│   ├── plans/             # Project plans not yet implemented
│   ├── decisions/         # Architecture decision records
│   ├── products/          # Product/feature documentation
│   ├── clients/           # Client work documents (private, gitignored)
│   ├── personal/          # Personal reference (private, gitignored)
│   └── reference/         # External API/spec references (private, gitignored)
└── projects/              # Archived project source code

Getting started

1. Copy templates to real files:

   cp AGENTS.md.tmpl AGENTS.md
   cp CLAUDE.md.tmpl CLAUDE.md
   cp agent-infra.md.tmpl agent-infra.md
   cp knowledge/INDEX.md.tmpl knowledge/INDEX.md
   cp knowledge/GRAPH.md.tmpl knowledge/GRAPH.md

2. Fill in your values — replace all <YOUR_...> placeholders in agent-infra.md with your actual bot tokens, host IPs, and credentials.

3. Point your AI agents at AGENTS.md — load it as a system prompt or project instruction file. Claude Code picks it up automatically if placed in the project root.

4. Start filing knowledge — follow the writing protocol in AGENTS.md. Agents will populate knowledge/ as they discover things.

How agents use it

Agents follow a lookup chain before asking the user:

1. Check the retrieval table in AGENTS.md for the right directory
2. Read knowledge/INDEX.md to find the relevant file
3. Read the file
4. Only ask the user if the above fails

When agents discover something new, they write it to the appropriate knowledge/ subdirectory and update INDEX.md.

Security

• agent-infra.md contains live credentials — it is gitignored. Only agent-infra.md.tmpl (with placeholders) is committed.
• knowledge/clients/, knowledge/personal/, and knowledge/reference/ are gitignored — add your sensitive content there freely.
• Run Gitleaks before pushing if you extend the whitelist.

Weekly maintenance

A lint agent runs weekly against knowledge/LINT-RULES.md to:

• Catch INDEX.md inconsistencies
• Find broken internal links
• Flag stale research notes
• Regenerate GRAPH.md monthly

Configure the schedule in your own cron or CI setup.