prism-mem

Every coding session leaves behind artifacts. Prism reads them and turns them into structured, reusable knowledge — automatically.

Prism is a post-session knowledge crystallizer for AI coding agents. It reads Claude Code session transcripts and git history, extracts semantic knowledge as triples, links them into a graph, and auto-regenerates the context files every agent reads at session start (CLAUDE.md, .cursorrules, AGENTS.md).

Quickstart

pip install prism-mem

# Configure your LLM provider (one-time setup)
prism config set provider anthropic
prism config set model claude-haiku-4-5-20251001
prism config set api-key <your-api-key>

# Run once after a coding session to extract knowledge and regenerate context files
prism crystallize --project /path/to/your/project

# Install the git hook so it runs automatically after every commit
prism hook install --project /path/to/your/project

# Add as an MCP server so agents can query the knowledge graph directly
claude mcp add prism -- prism serve --project /path/to/your/project

Configuration

Prism uses ~/.prism/config.toml to store your LLM provider settings. Set it up with:

prism config set provider anthropic        # or: openai, gemini, ollama, groq, mistral, ...
prism config set model claude-haiku-4-5-20251001
prism config set api-key <your-api-key>

prism config show                          # verify (api-key is masked)

The provider is validated against LiteLLM's provider list at set-time. Any provider LiteLLM supports works — Anthropic, OpenAI, Gemini, local Ollama models, etc. The model string is passed as provider/model to LiteLLM.

How it works

[git commit]
     │
     ▼
[1] INGEST
    Read ~/.claude/projects/<id>/<session>.jsonl  +  git diff HEAD~1 HEAD
     │
     ▼
[2] EXTRACT
    kg-gen → NetworkX graph of (subject, predicate, object) triples
     │
     ▼
[3] STORE + LINK
    Embed triples (sentence-transformers, local)
    Store in SQLite at ~/.prism/projects/<hash>/graph.db
    Connect related triples by cosine similarity (numpy)
    Flag contradicted facts as stale
     │
     ▼
[4] GENERATE
    Score triples (recency + confidence)
    Top 30 → LLM (via LiteLLM) → write CLAUDE.md / .cursorrules / AGENTS.md

Commands

MCP tools

When added as an MCP server, Prism exposes three tools to any agent:

Graph UI

prism ui opens a local browser UI at http://localhost:7823:

• /constitution — current CLAUDE.md with a Regenerate button
• /memory — searchable table of all triples (active + stale)
• /graph — interactive force-directed knowledge graph

Storage

All data is local. Nothing leaves your machine except calls to your configured LLM provider.

~/.prism/
├── config.toml             ← provider, model, api-key
└── projects/
    └── <project-hash>/
        └── graph.db        ← SQLite, one file per project

Requirements

• Python 3.13+
• Any LiteLLM-compatible LLM provider (configured via prism config set)

License

MIT