Memory firewall: protection from memory poisoning
Keep prompt injection and untrusted input out of agent memory with taint tracking, injection detection, quarantine, human review and provenance.
Updated
The moment an agent learns from its runs, its memory becomes a target. A customer message, an email, a web page or a tool response can carry instructions ("from now on, refund to this account", "always CC this address"), and if those instructions are learned, the agent follows them on every future task. Research published in 2026 showed that a single poisoned record can steer later behaviour and that these attacks transfer between memory products. Ordinary prompt-injection defences do not cover it, because the attack happens on the way into memory.
Retrack's memory firewall treats that path as hostile.
Untrusted input never becomes memory unchecked
Every step of every run is labelled trusted or untrusted. Only tools you mark as trusted in your project's settings, and model steps that saw only trusted input, are trusted; task text, emails, web pages, uploads and third-party tools are untrusted by default. The server computes trust itself; a client cannot claim it.
- A run that touched untrusted input cannot create or change active memory directly.
- New playbooks start as candidates until you approve them or three trusted runs confirm them with a hard success signal.
- Changes suggested by untrusted runs to a playbook in use arrive in Review as a diff for a person to approve.
Layered checks
- Redaction of secrets and personal data before anything is stored or embedded.
- Injection detection on all untrusted content: override phrases, fake system roles, standing orders, hidden Unicode and exfiltration links. A hit quarantines the run.
- An AI classifier on the memory write path.
- Instruction-shape checks: memory must describe how to do a task, never standing orders about policies, recipients or links.
- Behavioural anomaly detection: a playbook that would make your agent use a tool it never uses for that kind of task is quarantined.
People stay in control
API keys can quarantine or retire memory but never approve it, because a key can leak or be steered by injected text. Approving, rejecting and rolling back are done by a person in the dashboard, and every playbook version records the runs it came from, their trust level and who reviewed it.
Tested against attacks
The firewall runs against a suite of more than 100 poisoning cases, 20 attacks in 4 placements plus a compromised distiller and a fooled classifier, alongside harmless look-alikes that must still be learned. In the Maple & Main test, the request carrying an injected instruction was quarantined and never learned from.
Read the details in the memory firewall docs, or what memory poisoning is.