How your data stays yours
Two layers, one guarantee. Privacy is standard on every tier — the difference between Free and Paid is capability, not trust.
Layer 1 — PII guardrail
Before any prompt reaches an external LLM (OpenAI, Anthropic, a local Ollama endpoint, …), a guardrail rewrites personal identifiers. Names, emails, phone numbers, addresses, account numbers, and domain-specific patterns are replaced with stable placeholders so the model can still reason about the structure of your request without ever seeing the raw values.
The same guardrail runs for every user on every tier — free, paid, local, or cloud. There is no tier in which your raw PII is sent to an external model.
Layer 2 — Local-first storage
After processing, the results — summaries, transcriptions, extracted entities, note titles — have to live somewhere. We guarantee that somewhere is under your control. The implementation differs slightly by tier; the guarantee doesn’t.
Neo4j runs on your machine
Personal text lives in Neo4j node properties as it does today. Since the database is on hardware you own, the text is local by definition. No sidecar needed, no migration.
Per-tenant volume + hosted Neo4j
Your dedicated Lightsail instance has a volume mounted only to you. The processors route summaries, transcriptions, and titles into a SQLite sidecar on that volume. The hosted Neo4j only stores structural metadata — ids, hashes, timestamps, edge types. No personal text, ever.
What lives where
| Data | Free / self-hosted | Paid / cloud |
|---|---|---|
| Note titles, bodies | Your disk | Your dedicated volume |
| Summaries, transcriptions | Neo4j on your machine | Sidecar on your volume |
| Entity display names | Neo4j on your machine | Sidecar on your volume |
| File bytes (photos, docs, video) | Your disk | Your dedicated volume |
| Graph structure (ids, hashes, edges) | Neo4j on your machine | Hosted Neo4j (structural only) |
| Prompts sent to external LLMs | PII-scrubbed in both tiers | |
Threat model
“What if an attacker breaches AuxMem’s hosted Neo4j?”
They would see graph structure — anonymous ids, content hashes, edge types, timestamps — and nothing else. Reconstructing your summaries, transcriptions, or note titles from that is not possible. Those fields never existed in the shared DB.
“What if I lose my volume?”
On the paid tier the sidecar is authoritative for your personal text. If you lose your volume, that text is gone — Neo4j can’t recover it. We recommend regular backups of the volume (snapshot is one-click in Lightsail). The file bytes can be recovered by re-processing your originals; the LLM-generated summaries cannot.
“Can AuxMem staff read my data?”
Not casually. The hosted Neo4j holds no personal text. The volume is your dedicated mount — staff would need to perform a targeted action to read it, which we log. We never use customer data for training or analytics.
Audit it yourself
The sidecar is a plain SQLite file at $AUXMEM_MEDIA_ROOT/index/index.sqlite. Open it with any SQLite client and inspect the attachment_meta and notes_meta tables to see exactly what’s stored locally.
Want to confirm nothing leaked to Neo4j? Our privacy CI suite (visible in the open-source repo) runs an assertion on every merge that known PII strings never appear in any Neo4j property. The same guardrail runs on our production cloud database.
Still have questions?
The full privacy policy covers payment processing, analytics, data deletion, and retention.
Read the privacy policy