✦ Asterism

A local-first personal knowledge graph that thinks like a brain and looks like a constellation.

Every conversation you have with Claude leaves a trace. Asterism maps those traces into a living star map — the more you think about something, the brighter it glows. Stop thinking about it, and it fades into the dark.

What it looks like

How it works

• Hebbian learning — edges between concepts strengthen each time the LLM traverses them (weight += 0.2 per traversal). Concepts that aren't revisited accumulate session exposure time; after 3 hours of uninterrupted exposure without traversal they decay and vanish from the graph.
• You are the central node — your user node (Bidit) sits at the centre of the constellation at full brightness, always. It never decays.
• Local SQLite storage — the entire graph lives in ~/.asterism/asterism.db. No cloud, no sync, no accounts.
• LLM context injection — on every message, the top-N most relevant nodes and edges are injected into the Claude prompt as implicit context, letting the model answer with awareness of your past thinking.
• Triple extraction — each exchange is processed by a fast extraction model (local Ollama or Anthropic Haiku) that pulls (entity, relationship, entity) triples and writes them to the graph.

Quick Start

git clone https://github.com/biditdas18/asterism
cd asterism
chmod +x setup.sh && ./setup.sh

Then:

1. Run asterism init — guided setup wizard
2. Export your Claude data: claude.ai → Settings → Export Data → check email → download zip
3. asterism crawl --source claude --path path/to/conversations.json
4. Your constellation opens automatically

That's it. Your mind as a constellation.

Privacy

Your graph never leaves your machine. The only external calls are to the Anthropic API for Claude responses and (optionally) Haiku-powered triple extraction. The LLM only sees what you explicitly inject from your local graph — it has no access to the raw database. Delete ~/.asterism/ to get a clean slate. No telemetry, no analytics, no accounts.

Research

Asterism introduces several novel contributions to personal AI memory systems:

• Hebbian knowledge graph — edges strengthen on traversal, decay on disuse
• Session-based TTL — decay counts only active session time, not wall-clock time
• Graph-as-index — the graph acts as a weighted B-tree index over LLM flat memory
• Four biological mechanics — strengthening, pruning, pathway optimization, chain healing
• Causal chain topology — conversations organized by how ideas evolved, not just topic similarity

A formal research paper is in preparation. See RESEARCH.md for full details.

Built with

Layer	Tech
Storage	SQLite (~/.asterism/asterism.db)
Graph	NetworkX
Visualization	Vanilla JS force simulation (zero dependencies)
LLM	Anthropic SDK — claude-sonnet-4-6
Extraction	Ollama llama3.2:3b (local) or Anthropic Haiku (cloud)
UI	Streamlit
CLI	Click

Author

Bidit — github.com/biditdas18