Many vendors, many memories
ChatGPT, Claude, Copilot and in-house agents can each keep their own memory. A fact stored in one may be invisible to the others.
In development Design partner program
The flight recorder for what your AI agents know.
See which fact every agent used, who planted it, and fix it everywhere.
Sample record
2026-11-10 14:02 · fact v3 “Refund limit €500” → read by support agent for case 1847 → refund issued
01 / The problem
When a fact goes wrong, where do you look?
Memory can be poisoned
A single planted instruction can steer an assistant’s answers long after the original conversation.
No trail when it goes wrong
When an agent acts on a bad fact, teams can struggle to tell which version it used or who else read it.
In February 2026, Microsoft reported attempts by 31 companies across 14 industries to plant promotional instructions in AI assistants’ memory. Microsoft Security Blog
02 / How it works
One record across vendors.
is designed to sit between your agents and the memory they use. It records each read and write that passes through it.
-
Agents
ChatGPT, Claude, Copilot and in-house agents.
-
Records which fact changed, where it came from and which agents read it.
-
Any memory
Mem0, Zep, a vector database or
’s own store.
03 / Illustrative incident
One wrong fact, six steps.
Follow a bad refund limit from the moment it enters memory to the evidence an auditor can check.
Planted
An unapproved note raises the refund limit from €50 to €500.
13:58 · WRITE v3Sourced
The log ties the change to the note and the agent that saved it.
13:58 · SOURCERead
The support agent reads the €500 limit for case 1847.
14:02 · READ v3Acted
The case record shows a refund issued using that version.
14:02 · REFUNDRecalled
A reviewer withdraws v3 and restores €50 for every connected agent.
14:11 · RECALL v3Proven
An auditor verifies the signed record of the incident and correction.
14:18 · VERIFY
Sample correction
2026-11-10 14:11 · fact v3 recalled → “Refund limit €50” restored → earlier reads retained
04 / Design requirements
Built for the people who answer for incidents.
We’re building with security and compliance teams in regulated industries.
- Runs in your environment
- Keep deployment and audit records inside your organization’s boundary.
- Signed, tamper-evident log
- Give reviewers a way to detect changes to the audit record.
- Offline verification
- Give auditors an open-source verifier they can run without a connection to .
- Works with the assistants you already bought
- Keep your existing assistants and memory stores.
05 / Independence
Neutral by design.
sells no agents, models or memory. It can audit the tools you choose, and its own store is optional.
Our mission
Every fact an AI agent acts on should be traceable, correctable and provable.
06 / Design partners
We’re working with a small group of design partners.
If your team investigates AI incidents or has to explain them to an auditor, we’d like to hear how you work today.