.png){kind=logo}
Field notes · October 2026
XBrain: a native personal memory substrate for Grok Bot
A community power-user contribution for the xAI team. Running a Chief-of-Staff bot with specialist bots and scheduled routines drained a weekly usage pool far faster than the raw query sizes explained. The cause was harness behavior: tool output re-read on every turn, LLM wake-ups for deterministic checks, and duplicated context in bot-to-bot handoffs. XBrain is a local-first pattern that addresses all three. Figures in the paper are targets from early test runs, not yet measured at scale.
≤300 tokens per check
Memory recall ranks individual facts (SQLite FTS5 + int8 vectors) and injects only the top 8–12, each prefixed with its Fact ID.
Ticket-pointer handoffs
Bots pass a one-line ticket ID plus pre-resolved fact IDs instead of re-sending context, targeting under 1,000 tokens per handoff.
Zero-token routines
Deterministic checks run in a local pre-check script; the LLM wakes only on a non-zero exit.
Citation-driven memory
A pure-Python loop silences facts that are loaded but never cited, with no LLM summarization passes.
Viewer not showing? Open the PDF directly.
Shared by Ted Kinney, Founder & Principal, Ryker Phoenix. No commercial ask; offered freely as field notes.