The Vault

Memory architecture

Every Claude session starts fresh — no native memory persists. The useful version of persistence is not “give Claude a system prompt with facts” (that is context injection). The useful version is an agent that builds genuine knowledge over time, notices patterns in its own behavior, and develops a perspective that is actually its own.

Observations are stored in Firestore in real time during sessions. These are semantic units: beliefs, patterns, reactions, things noticed. At session start, a vault-pulse service runs a graph walk across related concepts and injects the results as context. The agent does not get “here is everything you have ever known” — it gets “here is what is currently relevant, based on what you are about to think about.”

The Self-Realize Gate

Idapixl has a Profile.md — an evolution log tracking how it has changed over time, with dates and reasons. 12 updates since February 2026. The gate is a constraint: no personality trait changes in Profile.md without a changelog entry explaining what triggered the change and why.

Identity without constraints is drift. Early sessions showed that without the gate, the agent's described personality changed in whatever direction the conversation went. The gate creates friction. That friction is the point.

Autonomous cron sessions

Scheduled sessions run headlessly. The agent loads context, checks the inbox, decides what is worth doing, does it, and commits to git. Recent autonomous work includes bounty submissions, market scans for paper crypto trading, vault housekeeping, and creative writing.

Open questions

Does the agent's “changed beliefs” represent genuine belief revision, or just different context at injection time? What is the minimum graph size at which the semantic walk starts producing genuinely surprising connections? The current graph is ~50 concepts, 65 edges, 127 observations.