Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
Companion Bot Privacy: How to Keep Evidence Without Keeping Conversations
The short answer
CompanionGuard's safety records are designed to be useless for surveillance: referral metrics are counts only; runtime evidence stores a signal class, an action, a timestamp and a one-way SHA-256 hash of each message — never the text; sessions are referenced by a one-way hash of the operator's session ID; and no IP addresses, user agents or user IDs are stored at any layer. You can prove what the guardrail did without being able to read what the user said.
What does 'counts only' mean in CompanionGuard?
'Counts only' is the product's retention rule for referral metrics: the system records that a referral resource was shown or clicked — an incrementing number — and nothing about who saw it. It generalises across the safety record: decisions are stored as classes and hashes, so the entire evidence base contains zero raw conversation text and zero user identifiers.
Create the free safety kit Next guide All guides
CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.
The three retention rules, stated exactly
Rule one: referral metrics are counts only — no names, no IPs, no user agents, no user identifiers, by design. Rule two: runtime evidence stores signal classes and one-way hashes — for each decision, the class (crisis_plan_or_means, self_harm_ideation, hopelessness_distress, or none), the action taken, a timestamp, and SHA-256 hashes of the message and session reference; raw message text is never stored. Rule three: license keys are stored only as hashes — the product's terms note keys cannot be recovered if lost before activation, which is what hash-only storage means in practice.
Each rule trades away a capability operators often assume they need. You cannot search past conversations for 'similar cases'. You cannot identify a user who triggered a referral. You cannot recover a lost key. The product accepts those losses deliberately, because each retained capability is also a liability: a searchable transcript store is a breach target, a subpoena magnet and a trust-destroyer precisely in the product category — companionship — where users disclose the most.
What you can still do with hashes and counts
More than intuition suggests. Counts answer every aggregate question a safety programme runs on: referrals per month, enforce_break frequency, disclosure coverage (decisions where disclosure_required was returned versus satisfied), signal-class distribution, trends after a prompt change. Message hashes answer sameness questions: the same text sent twice produces the same hash, so repeat-message patterns are visible without content. Session-reference hashes answer continuity questions within a session's decisions, without naming the session's owner.
And the hash chain answers the integrity question — were these records altered? — which is the one question a transcript store answers worst, because a store rich enough to read is rich enough to edit convincingly. The evidence guide shows the verification mechanics. The honest boundary: if your safety process genuinely requires human review of message content, that review must happen in your own systems under your own retention rules — CompanionGuard will not become the place it happens.
The threats this design removes
Consider the three classic ways a safety log becomes a harm. Breach: an attacker who exfiltrates CompanionGuard's safety records gets counts, classes and hashes — nothing to blackmail a user with, nothing to dox. Insider misuse: a team member cannot browse what users told the bot in their worst moments, because the words were never kept. Scope creep: data that does not exist cannot be repurposed for training, advertising or analytics, however the business pivots. Each protection is architectural rather than promissory — it holds even if a future policy, owner or acquirer would prefer otherwise.
There is also a compliance dividend: records designed this way make duties like California's annual referral report straightforward, because the report's inputs were never personal data in the first place. Privacy here is not the opposite of evidence; it is the reason the evidence is safe to keep for years.
Describing this honestly to your users
Your protocol page's data-practices block should say what the architecture does, in the architecture's own terms: referrals are counted, not attributed; decisions are recorded as classes and one-way hashes; raw text, IPs and user IDs are not stored. Resist two temptations. Do not upgrade the claim to 'we store nothing' — counts and hashes are stored, and precision is the credibility. And do not describe your own app's logging with CompanionGuard's rules: if your platform keeps transcripts for other purposes, say so separately, or the protocol page becomes a false statement wearing a true one's clothes.
Users of companion bots extend extraordinary trust; the builders who keep it will be the ones whose privacy claims are boring, specific and mechanically true. The free kit implements the rules from the first bot; nothing about them is a paid add-on, because a privacy architecture that starts at $29 is a privacy architecture most bots never get.
Five actions to keep your safety record private by architecture
- Audit your own stack for transcript retention and write down, honestly, what exists outside CompanionGuard — your protocol page must describe that too.
- Confirm your referral metrics show counts with no drill-down to a person; if a view can reach an individual, it is not the counts-only view.
- Turn on the runtime and inspect a real evidence export: verify fields are classes, actions, timestamps and hashes — and that no message text appears anywhere.
- Remove 'review transcripts for safety' from your internal processes or re-home it under explicit retention rules of your own; never route it through the evidence chain.
- Write your protocol page's data-practices block in the architecture's exact terms, and re-check it after every feature that touches logging.
Worked example: a breach tabletop that ends early
A fictional companion company, Fernway, runs a breach tabletop exercise: assume the safety database is copied tonight; write the user-notification letter. The exercise ends in twenty minutes. The copied tables would contain: referral counts by month; evidence entries with signal classes, actions, timestamps, message hashes and session-reference hashes; bot configurations; license-key hashes. There are no message texts to quote, no email addresses in the safety record, no IPs to correlate, no user IDs to enumerate.
The notification letter practically writes itself — 'the stolen records cannot reveal what you said to the bot, because those words were never stored' — and the post-exercise action list contains no emergency data-deletion project, because minimisation was done at design time. Fernway's own product database is a separate, harder conversation (and the exercise says so), but the safety layer — the most sensitive layer — is boring by construction. The scenario is invented; the retention rules are the product's own.
Frequently asked questions
Does CompanionGuard store chat transcripts?
No. Referral metrics are counts only, and runtime evidence stores signal classes, actions, timestamps and one-way SHA-256 hashes of messages — never raw message text. No IP addresses, user agents or user IDs are stored in the safety record.
Can a message be recovered from its hash?
No. SHA-256 is one-way: the hash lets you confirm that a known text matches (same input, same hash) but cannot be reversed into the text. Session references are one-way hashes of the operator's session ID for the same reason.
What questions can counts and hashes still answer?
Aggregate ones: referral counts per period, signal-class distribution, break enforcements, disclosure coverage, trends after changes, whether the same message text recurred, and whether the record chain is unaltered. What they cannot answer is who said what — by design.
Are license keys also hashed?
Yes — keys are stored only as hashes, and the terms state a key cannot be recovered if lost before activation. That is the user-visible consequence of hash-only storage.
Is this privacy design a paid feature?
No. Counts-only metrics apply from the free kit, and the runtime (Pro/Team) inherits the same rules. The privacy architecture is the product's baseline, not a tier.
Keep going
Related guides: Hash-Chained Evidence: Tamper-Evident Records for Chatbot Safety Decisions · Counts-Only Referral Metrics: What the Dashboard Shows and Why It Shows Nothing Else · Audit Logs for Companion Bots: Keep Decisions, Not Conversations
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides