A Wedding Website Becomes a Test of AI Memory

Summary

While building a wedding-invitation website, the author found that a coding LLM either asked extensive clarifying questions or spent costly tokens repeatedly exploring the same project. To preserve useful discoveries across sessions, they built `@pureooze/pi-knowledge-graph`, an extension for Pi, a lightweight coding-agent harness. The extension stores entities, claims, and evidence linking facts to files, commands, user statements, URLs, or Pi sessions. Its `knowledge_search`, `knowledge_get`, and `knowledge_maintain` tools let an agent retrieve project architecture and prior decisions, inspect supporting sources, and add, update, or remove records. When a fact changes, a replacement claim supersedes the old one while preserving its history and evidence, allowing the agent to distinguish current information from stale context. The database is shared across Pi sessions and stored locally under `~/.pi/agent/knowledge-graph/`; retrieval uses SQLite full-text search and graph connections without embeddings or additional model calls. The author warns that durable memory can preserve incorrect information, and changes made through the maintenance tool are not undone by conversation rewinds or branches. The package requires Pi 0.84.0 or newer and Node.js 24.14.1 or newer, and includes commands for status, export, and record deletion.