You're trusting us with your memories.
Lodestone holds what you tell your AI: your work, your decisions, sometimes your life. We treat that as a responsibility, not a feature. So every protection on this page names the mechanism behind it, not just a promise, and we keep what works today separate from what we're still building.
Built into Lodestone today
In Lodestone itself, whether you host it yourself or we host it for you.
Secrets are redacted before anything is written
Private keys, API tokens (Anthropic, OpenAI, GitHub, AWS, Slack, Stripe, Google and more), JWTs, bearer tokens and passwords in URLs are replaced with [REDACTED:type] on your machine, before a transcript is archived. Self-hosters can add their own patterns.
A token per device, scoped and revocable
Each laptop, phone or teammate gets its own token, limited to the namespaces you choose, read-only or read-write. Revoking one ends it and every app connection made with it, at once. Tokens are stored hashed, never in git.
Standard OAuth for claude.ai and other apps
Remote apps connect through OAuth 2.1 with PKCE and rotating refresh tokens. Redirects are limited to an allowlist, refused attempts are logged, and only hashes of secrets are stored.
Nothing is shared unless you share it
Only you can share a note or a folder, from the web UI or the CLI; no agent tool can. Shares are read-only, can expire, and stop on the next request after you revoke them. People you share with cannot pass your notes on.
Memory is treated as data, not instructions
Everything an agent reads back from Lodestone is wrapped and labelled as retrieved data, so text stored in a note cannot pose as instructions to the agent reading it.
Nothing is lost silently
Every change is a git commit, and notes are retracted rather than deleted. If storage or backups are at risk, the server refuses new writes and says why, instead of keeping data where the next deploy would wipe it.
It will not go public without a lock
The server refuses to listen on a public address unless access tokens are set.
Your memory is plain files you own
Markdown and JSON in a git repo. You can read every byte, keep your own copies and leave at any time.
What we commit to when we host it
Cloud is in early access. These are the guarantees it launches with.
HTTPS only
Every connection is encrypted in transit with TLS. HSTS tells browsers never to use plain HTTP, and plain HTTP requests are refused, never served.
Secrets and backups encrypted
GitHub tokens and any API key you give us are encrypted (AES-256-GCM) and bound to your account. Nightly backups are encrypted before they leave the server. Your memory itself is not yet encrypted with its own key: that comes in Phase 1, with a key per account.
Nobody reads your memory
Nobody at Lodestone reads your memory as part of running the service. Until Phase 1, the person who runs the server could technically read it, and we say so in your Lodestone. If we ever need to look to fix something, we ask you first, or tell you afterwards if it was urgent.
Every account is isolated
Each account has its own folder and its own git repo. Automated tests try every endpoint and agent tool with one account's credentials against another's data, and must be refused.
Leave with everything, any time
Download your notes as plain Markdown or the whole git repo with its history, or keep a live mirror in your own GitHub. Delete your account and your data, backups and keys are gone within 30 days.
Your model key stays yours
If you bring your own model API key, it is encrypted, used only for your own distillation, never logged, and you can delete it at any time.
Self-hosting? Keep it on HTTPS
When you run the server yourself, put it behind HTTPS: hosts such as Railway do this for you on their domains. Lodestone's protocol requires HTTPS for servers that share with each other, and claude.ai connects only to HTTPS addresses.
Our responsibility
When safety or your control over your data conflicts with convenience or cost, we choose safety and tell you why. If we ever change how your memories are stored, protected or shared, we'll tell you before it happens, not after.
Found a problem?
Please report security issues privately through our contact form. Reports come straight to us and are never published. We'll reply, fix the issue, and credit you if you'd like.