ADAM is operated by [LEGAL ENTITY NAME], contact [CONTACT EMAIL]. Where ADAM is self-hosted on your own infrastructure ("Sovereign"), you are the operator and controller of all data it holds; this policy then describes the software's behaviour rather than any processing by us.
Only from machines that have recorded consent:
| Category | Detail |
|---|---|
| Terminal activity | Commands you run and their output, truncated to ~4,000 characters and secret-scrubbed before leaving your machine. |
| Development context | Git commit messages, branch and repository names; paths of files recently worked on. |
| AI tool sessions | Claude Code / Gemini CLI session transcripts, distilled into a short question-and-answer "lesson" rather than stored verbatim. |
| Project code & docs | Only for directories you explicitly index. Never scanned automatically. |
| Machine metadata | Hostname, OS, client version, a hash of the client code, and check-in timestamps — for fleet management. |
| Usage records | One row per question asked: timestamp, whether it was answered from memory, latency, answer length. Used for your own weekly value report. |
ADAM does not keystroke-log, does not screen-capture, does not read arbitrary files outside directories you index, and does not capture passwords typed at prompts (those are not echoed to the terminal).
Collection is off until consent is recorded on that specific machine. The consent ledger is append-only and the check fails closed — an unknown machine, or an unreadable ledger, results in refusal to collect rather than collection by default.
~/.adam/paused, or run the supplied pause command. Collection stops immediately.ADAM_NO_TRANSCRIPT=1 to exclude a session.Before data leaves your machine, and again on arrival, ADAM masks common secrets:
API keys, bearer tokens, JWTs, private key blocks, database connection strings,
KEY=value secrets, email addresses, IP addresses and high-entropy blobs.
The rules are deliberately over-broad — masking a harmless UUID is preferred to leaking a token.
This is best-effort, not a guarantee. No pattern-based scrubber catches everything. Please continue to avoid pasting production credentials or customer data into your terminal.
When ADAM cannot answer from memory, it forwards the question — not your corpus — to an AI provider (for example Google Gemini, Groq, OpenAI, Anthropic, OpenRouter or xAI). On Bring-Your-Own-Key plans these calls run on your provider accounts under those providers' terms. Retrieved context from your own corpus may be included in the prompt where it is needed to answer.
Stored knowledge is tiered. The lower tiers are eligible for automatic expiry on the schedule below:
| Tier | Retention |
|---|---|
| Working | ~24 hours |
| Scratch (bulk harvested) | ~30 days if never retrieved |
| Episodic / Curated (validated answers) | Retained until you delete it |
Whether automatic expiry is switched on depends on the deployment, and on this one it is currently OFF — data in the lower tiers is retained until it is deleted on request (see below) rather than ageing out on its own. We would rather say this plainly than publish a schedule we do not apply. An entry that has been retrieved at least once is never expired automatically in any case.
Self-service. Authenticated with your tenant key you can export
everything held for you at any time (GET /v1/saas/export, JSON or JSONL),
and erase it irreversibly (DELETE /v1/saas/data, confirming your tenant id).
Erasure covers your knowledge, usage records, unanswered-question queue, account row,
stored provider-key references and search index, and revokes your API key. Provider key
values are never included in an export.
You may also contact [CONTACT EMAIL]; requests are actioned within 30 days.
ADAM keeps a general pool of lessons used to improve the product for everyone. Your team contributes nothing to it unless an owner of your team switches that on, and switching it on records who made the decision. It can be switched off again at any time.
Opting in does not publish anything by itself. Every candidate lesson passes through two further steps:
Only lessons that have been validated and scored highly are ever candidates. Raw telemetry, low-quality entries and anything belonging to another team are never offered.
Depending on where you live, you may have rights to access, correct, export, delete, or restrict processing of your personal data, and to complain to a supervisory authority. Exercise them via [CONTACT EMAIL].
Where the hosted brain is backed up, the archive is encrypted with a passphrase held by the operator and contains the knowledge database and the encrypted secrets store. On this deployment a snapshot is taken nightly, kept for about 7 days on the server, and a copy is held for about 30 days on operator-controlled storage in a different location so that losing the server does not lose the data. A backup taken before you delete your data will still contain it until that backup ages out, so erasure under §7 is immediate in the live system and completes in backups within roughly 30 days. The hosting provider additionally takes its own whole-server snapshots on its own schedule.
Authentication by scoped API key; administrative and code-execution endpoints are refused over the public network entirely; per-identity rate limits; secrets encrypted at rest. No system is perfectly secure, and we do not claim otherwise.
Material changes will be announced before taking effect, and continued collection after a material change requires renewed consent.