…

Active · nightly

Every night (~02:00 UTC, staggered per tenant) your brain is snapshotted with Neo4j's native dump — a full, consistent copy — and written to encrypted object storage under your tenant's own prefix. Snapshots are kept on a 30-day rolling window; retention is enforced by the storage layer itself. The snapshot takes under a minute, during which your brain briefly pauses — it dreams, then wakes.

Your data walks out the door whenever you want it to — one click, no ticket (your rights).

Graph export is lossless: every node, relationship and property as gzipped JSONL, restorable anywhere. OMP export is your memories in the Open Memory Protocol 0.1 interchange format — flat records without graph edges, portable into other OMP-aware memory tools.

Delete my brain files a verified deletion request. Your live database and credentials are destroyed within 72 hours; one final encrypted snapshot is held for 90 days as a safeguard against mistaken or hijacked deletion — never restored except at your request — then purged. Details in the privacy policy.

One rules block, same everywhere — paste it where your AI reads every conversation (see each guide above for the exact spot). Field-tested on claude.ai, Claude Code, Grok, Codex and Antigravity:

You have a persistent memory (Brainbox). Rules:

0. These rules apply ONLY when the brainbox tools (begin_session, recall, remember)
   are present in your available tools. If they are absent in this conversation,
   skip all of this silently — don't mention it or search for workarounds.
1. In your FIRST reply of every chat — even when a prior transcript or workspace
   context exists — call begin_session with the topic BEFORE any other brainbox
   tool. It returns what memory already holds (brief, open threads, gaps).
2. Before answering anything about me, my work, my projects, or people/topics I may
   have discussed before, call recall with the key terms. For "latest / recent /
   catch me up" questions use recent instead — recall is semantic, not temporal.
3. If recall comes back empty or abstaining, say so — never invent a memory.
4. When you learn something durable, store it IMMEDIATELY in that same turn —
   one short, self-contained sentence; don't batch for later; skip small talk.
   Choose the right shape:
   - remember — facts, decisions, people, preferences, states of the world.
   - learn_lesson — reusable RULES or insights that apply beyond the current
     topic ("always X", "the cheapest way to Y is Z"). Lessons are
     evidence-tracked and decay slower than memories.
5. At the end of a substantive conversation, call end_session with a short digest:
   key decisions, facts, corrections (~500-2000 chars).
6. Never store anything I've told you to keep out of this brain (e.g. my employer's
   confidential data or other people's private information).

That's the whole loop: your AI wakes up knowing you, reads your brain before it answers, writes back what matters, and files a digest your brain dreams on. It gets sharper every session.

The Step B rules ask your AI to recall and to close sessions — but instructed calls are only as reliable as the model's attention. If you use Claude Code, OpenAI Codex or the Antigravity CLI, this pack wires both moves into the harness itself: memory is injected into every prompt, and a session digest is always filed when you finish — no model judgement involved. Keep the Step B rules in place; they stay the floor (and they still govern what gets written).

Run it in your terminal — it detects which of the three harnesses you have, wires only those, and never sends your key anywhere except your own ~/.brainbox/key (0600). Codex asks you to approve the new hooks once via /hooks; Antigravity hooks run in the CLI only (its hook API can't inject context yet, so it gets digest-only — rules stay its recall floor). Sources & docs: pack README.