A template repository for building persistent organizational memory with an agent-maintained Markdown wiki.
The repository is not the agent runtime. It is the agent's durable memory: source configurations, raw evidence, reusable ingestion tools, and a living wiki inspired by Karpathy's LLM Wiki pattern.
- A wiki-centric structure for people, teams, projects, repositories, systems, decisions, and tools.
- GitHub ingestion scripts that preserve raw evidence before synthesis.
- A local BM25 search layer with optional sentence-transformer embeddings.
- Agent skills for the Observe → Ingest → Synthesize → PR loop.
- A clean starting point with no organization-specific data.
See docs/GETTING_STARTED.md for the full onboarding flow.
- Create a repo from this template.
- Edit
mini-agi.yamlwith your organization name and operating preferences. - Edit
data/sources/github-watchlist.jsonwith the GitHub orgs/repos you want to observe. - Authenticate GitHub CLI:
gh auth login- Smoke-test ingestion:
node tools/ingestion/github-daily-ingest.js --dry-run --limit 3- Start writing/synthesizing the wiki from evidence:
Observe → Ingest raw evidence → Synthesize wiki pages → Cross-link → Open PR → Repeat
.agents/ reusable agent skills
data/ source configs, raw evidence, generated search artifacts
docs/ reference docs, including the LLM Wiki pattern
tools/ ingestion and search tools
wiki/ living Markdown knowledge base
AGENTS.md operating instructions for agents working in this repo
mini-agi.yaml organization-level configuration
- Getting started
- Configuration
- Operating loop
- Ingestion pipeline
- Privacy and data handling
- Example walkthrough
BM25 works immediately and has no npm dependencies:
tools/search/build.sh
node tools/search/search.js "your query"Optional vector search:
pip install sentence-transformers numpy
python3 tools/search/embed.py corpus --model sentence-transformers/all-MiniLM-L6-v2 --normalize
node tools/search/search.js "your semantic query"This template is public-safe, but your generated repository may not be. Treat data/raw/, private wiki pages, and source credentials as sensitive unless you intentionally publish them.