AI Experiments

A collection of small, practical AI projects exploring local-first assistants, agent pipelines, retrieval, personal knowledge systems, and MCP servers. Each project is self-contained and has its own README with setup and usage details.

Projects

Project	What it does	Main ideas
daily-digest-agent	Builds a personalized daily Markdown digest from RSS, Hacker News, and Reddit.	Local or API-backed LLM scoring, summarization, deduplication, configurable sources.
llm-wiki	Maintains a persistent LLM-assisted wiki from source documents and saved queries.	Compounding knowledge base, source ingestion, Obsidian-friendly wiki pages, Anthropic or Ollama backend.
local-rag-assistant	Lets you chat with a local Notion export and extract structured timelines from it.	Local RAG, ChromaDB, SQLite, DuckDB analytics, Streamlit UI, Ollama models.
multi-agent-research-assistant	Produces cited research reports from a topic using planner, searcher, and writer agents.	Async multi-agent workflow, web search, provider abstraction, CLI and Streamlit interfaces.
tiny-memory-mcp	Provides durable local memory to MCP-compatible AI clients.	Model Context Protocol, SQLite-backed memory, stdio and Streamable HTTP transports, Docker/ngrok deployment.

Project Summaries

Daily Digest Agent

daily-digest-agent fetches new items from configured sources, filters out previously seen links, scores each item against your interests, and renders a clean Markdown digest. It can use Ollama for a fully local workflow or Groq for a hosted free-tier option.

Start here:

cd daily-digest-agent
pip install -r requirements.txt
python main.py

LLM Wiki

llm-wiki is based on Andrej Karpathy's persistent wiki pattern. Instead of getting one-off answers from a chatbot, you add sources to a durable markdown wiki that the LLM updates over time. Ingested documents become summaries, concept pages, index updates, and activity log entries; useful query answers can also be saved back into the wiki.

Start here:

cd llm-wiki
pip install -r requirements.txt
python scripts/ingest.py sources/my-article.md
python scripts/query.py "What are the main themes?"

Local RAG Assistant

local-rag-assistant is a private Notion knowledge assistant. It ingests a Notion Markdown/CSV export, chunks and embeds the content locally with Ollama, stores vectors in ChromaDB, extracts structured events into SQLite, and exposes chat plus timeline workflows through Streamlit.

Start here:

cd local-rag-assistant
pip install -r requirements.txt
cp config.example.yaml config.yaml
python scripts/run_ingest.py
streamlit run src/ui/app.py

Multi-Agent Research Assistant

multi-agent-research-assistant turns a research topic into a cited Markdown report. A planner decomposes the topic into sub-questions, search agents run those questions in parallel, and a writer synthesizes the findings. It supports Groq, Ollama, and OpenAI-style providers, with DuckDuckGo search available when Serper is not configured.

Start here:

cd multi-agent-research-assistant
pip install -r requirements.txt
python main.py "The current state of fusion energy"

Tiny Memory MCP

tiny-memory-mcp is a dependency-free Python MCP server that gives an AI client durable local memory. It stores memories in SQLite and exposes tools to remember, recall, list, fetch, search, and forget records. It can run over stdio for local MCP hosts or over Streamable HTTP for remote connector-style usage.

Start here:

cd tiny-memory-mcp
python3 server.py
python3 tests/smoke.py

Common Themes

• Local-first AI: several projects can run with Ollama so private data stays on the machine.
• Small, inspectable systems: most projects avoid large agent frameworks so the core mechanics are visible in plain Python.
• Persistent memory and knowledge: projects explore durable stores such as Markdown, SQLite, ChromaDB, and DuckDB.
• Agent workflows: the repo includes examples of single-purpose agents, multi-agent pipelines, tool use, and MCP integration.

Suggested Reading Order

1. Start with daily-digest-agent for a simple end-to-end local agent.
2. Read multi-agent-research-assistant to see planner/searcher/writer orchestration.
3. Explore local-rag-assistant for a fuller retrieval and analytics pipeline.
4. Try llm-wiki for a markdown-native personal knowledge workflow.
5. Finish with tiny-memory-mcp to understand how AI clients can use external memory through MCP.

Notes

• Each project has its own dependencies and setup steps.
• Generated data, local databases, API keys, and private exports should stay out of git.
• Check each project's README before running it; several require local services such as Ollama, Streamlit, Docker, or API keys.