Memory audit · v0.1 preview
Your AI assistants remember more than you think.
Smelten audits the memories stored inside Claude, ChatGPT, Copilot, and Gemini across your organization. It surfaces stale facts, orphaned memories from departed employees, and the quiet signatures of memory poisoning.
The problem
Assistant memory is instrumentation nobody reads.
01 / ACCUMULATION
Silent accumulation
Enterprise AI assistants accept facts from every session. Most users cannot open the store to see what is there. The contents only grow, and the assistant keeps answering with whatever is inside.
02 / DEPARTURE
Orphaned after departure
An employee leaves. Their assistant memories stay. When a colleague asks a related question, those memories still answer. Nobody marked them for removal, because nobody saw them.
03 / SIGNATURE
Poisoning leaves a signature
A prompt injection plants a fact that looks ordinary. Without instrumentation, it reads the same as anything else in memory. The signature is there, faint, and only visible under inspection.
What Smelten inspects
Six classes of memory. Six questions. One report per workspace.
- Stored user facts
- age · reference count · newer contradicting signal
- stale_fact
- Project memories
- owner employment status · shared scope
- orphaned_memory
- Preference flags
- drift vs observed behavior
- divergence_score
- Instruction memories
- injection signature · credential-like patterns
- poisoned_memory
- Cross-chat references
- leak potential across workspaces
- scope_violation
- Memory totals
- aggregates per workspace · per assistant
- audit_summary
Cross-section
A thin slice of a workspace's memory, viewed through instrumentation.
Each row is a class. Each cell is one memory. Flagged items sit in amber.
audit / claude · workspace: operations · 1,345 memories
61 flagged · 1,284 ordinary
How it works
Four steps from connection to approved purge.
- 01
Connect
Read-only OAuth to each assistant's memory store. No writes, ever. Nothing leaves your region. - 02
Scan
Every memory item classified against the Smelten signature library. Stale, orphaned, poisoned, or ordinary. - 03
Review
A plain-English audit report with the finding that led to each classification. - 04
Purge
You approve what leaves. Smelten never removes a memory without your sign-off.
Trust
What we touch. What we retain. What we cannot do.
- Read-only, always
- We do not write, modify, or delete memories. Purges happen only after you approve each one.
- Metadata over content
- Memory text is hashed and summarized. Full-text retention is opt-in, per workspace.
- Your region, your data
- Findings return to you. Nothing is retained on our side beyond audit metadata.
- SOC 2 path in progress
- Type I target: Q4 2026. Type II follows. We will not claim certification we do not hold.
Pricing
One tier while we build. Enterprise when you need it.
Starter
- Up to 3 connected assistants
- Monthly audit, delivered as report
- Metadata-only retention by default
- Email support, direct line to the team
Enterprise
- All connected assistants, all workspaces
- Continuous audit with change alerts
- SSO, custom retention, regional residency
- Shared channel with our team
Request audit
Start with one workspace.
Early access is free while we build v0.1. You receive the first audit and a direct line to the team.