Privacy Policy & Data Usage

Last updated: September 2026

1. What We Collect

DGUI-HyperMem collects only the minimum data needed to operate the memory service:

• Email address — used to issue and manage your API token
• Usage data — anonymized context categories (e.g., "memory_add", "memory_search") for quota tracking and service improvement
• Memory content — the memories you explicitly store via the MCP add tool. This data is stored in your D1 database and is never shared externally.

2. Training & Data Sharing

By default, your usage context (not your memory content) may be aggregated for model improvement. You can control this:

• Train with everyone ON (default) — anonymized usage patterns are used to improve the JEV reasoning layer
• Train with everyone OFF — no data from your token is used for training. All processing stays local.

Toggle this setting anytime via the Admin Dashboard — your admin can view and change your preference.

3. Secrets & Content Safety

Your memories are stored exactly as you wrote them. We do not rewrite your content, and a secret in a memory stays in that memory — it is your record, and altering it silently would be worse than the risk.

What we do protect is the training corpus. When a JEV decision is queued for dataset export, the row is passed through a redaction pass first, before it is ever written to the queue. Anything matching a known credential format (Stripe, GitHub, HuggingFace, Cloudflare, AWS, Google, Slack, OpenAI keys; webhook secrets; JWTs; bearer headers) or a keyword assignment such as passkey = … is replaced with [REDACTED:<rule>] in the training row only.

Two honest limits. Redaction is pattern-based, so a secret that is a single lowercase dictionary word with no digits, no format prefix and no surrounding keyword will pass through. And if the redaction pass itself errors, the row is dropped rather than exported unvetted. If you are storing something that must never leave your account, the reliable control is turning training off for your token — see section 2 — rather than relying on the scanner to guess.

4. Data Storage & Retention

• Memories are stored until you explicitly delete them via the forget tool
• Usage logs are retained for 90 days, then aggregated and anonymized
• You can request full data deletion by contacting the admin

5. Third-Party Services

DGUI-HyperMem uses:

• Cloudflare Workers — serverless compute (D1 database, Vectorize, Workers AI)
• TypeSafe AI — Jev / System One model for memory classification (only memory type/salience data, never raw content without your consent)
• HuggingFace — optional dataset sync (only if train_with_all is enabled)

6. Self-Hosting

All data stays within your Cloudflare account. DGUI-HyperMem is open-source under MIT license. You can self-host at any time — see github.com/ctaxnagomi/dgui-hypermem for deployment instructions.

7. Contact

For privacy inquiries, reach out via the DeckerGUI project page.

DGUI-HyperMem — DeckerGUI Project. MIT License.