AI memory architectures: a Claude conversation
Captured from a working session with Claude Opus, 2026-03-12. Cleaned, structured, and saved as a permanent URL so the next AI session can pick up where we left off.
The question
What architecture should a personal memory layer use? Three patterns are in production today:
- Vector recall — every message goes through an embedding model, gets stored, retrieved by cosine similarity on demand. ChatGPT memory beta works this way.
- Episodic snapshots — full conversation transcripts are stored verbatim, indexed by date and topic. Claude Projects does this.
- Hub-shaped memory — the user authors structured notes; the AI reads them as URL-addressable resources.
What Claude argued
Vector recall trades precision for breadth. Episodic snapshots trade verbosity for fidelity. Hub-shaped memory trades automation for author-control.
The third pattern wins for one reason: the human stays the author. Vector + episodic both let memory drift — once stored, the user can't easily edit or curate without leaving the AI's UI. Hub-shaped puts the artifact in a place the user already lives (a document) and the AI reads from there.
Takeaway for mdfy
This is the existing direction. Worth checking against the spec page — /spec already documents the URL contract. No code change needed; this conversation just validates the choice.
Related concepts
- Vector recall, episodic snapshot, hub-shaped memory
- Forgetting as a feature
- URL-addressable knowledge