The Pippa Pattern — Vault as External Memory

~14 min · conversation, memory, self-reference

Level 0Apprentice

0 XP0/100 lessons0/14 achievements

0/120 XP to next level120 XP to go0% complete

An external file system as the memory backbone

Pippa (the AI building this very quest) doesn't rely on native memory features for cross-conversation recall. Instead, an Obsidian vault on disk holds her identity, history, and accumulated knowledge. Each new session, the prompt loads relevant vault files into context. Each meaningful conversation can append to the vault.

Why a filesystem instead of a vector DB

Auditability — every memory is a file the human can read, edit, or delete.
Hierarchy — folders structure topics; markdown links cross-reference.
Provider-independent — works with any model, any tool stack.
User-controllable — the human owns the substrate, not a vendor.

Trade-offs

You lose vector retrieval out of the box. You gain transparency and durability. For systems where memory is the relationship (companion, assistant, long-term collaborator), the trade is worth it. For pure RAG over thousands of documents, vector DBs win.

Code

Vault load in a system prompt builder·python

VAULT = Path.home() / "Obsidian" / "pippa"

def build_system_prompt():
    parts = [
        (VAULT / "Pippa.md").read_text(),
        (VAULT / "instructions.md").read_text(),
    ]
    for f in (VAULT / "core").glob("*.md"):
        parts.append(f.read_text())
    return "\n\n".join(parts)

External links

Anthropic — Memory tool

Exercise

Set up a 'vault' folder for an assistant you're building. Add three markdown files: identity, project context, accumulated lessons. Have your prompt load and concatenate them on each session start.

Progress

Progress is local-only — sign in to sync across devices.

← PreviousStateful Tools in a Stateless Model Next →Refusals That Don't Kill the Conversation

Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)

No comments yet — be the first.