Instrumentation, paired
A 60–90 minute kickoff call where we wire ferryte.instrument() into your agent together. One line; no agent-code changes. You keep the instrumentation forever.
Fixed price · 48-hour turnaround · runs in your infra
Your agent’s memory has bugs
you can’t see. We’ll show you every one.
The Agent Memory Audit: we instrument your agent, run the full forgetting battery against your real memory backend, trace your worst wrong answers to the exact memories that caused them — and hand you the evidence, the blind-spot map, and the fix list. In 48 hours, for a fixed $500.
What you get
Five deliverables. One fixed price.
The full forgetting battery
All five structural scenarios against your real backend: source-revocation, cross-tenant-isolation, stale-fact, memory-poisoning, and mosaic reconstruction. Both retrieval results and raw store contents are checked.
Root cause for your worst answers
Bring up to three wrong, stale, or leaked answers. ferryte why traces each one to the memory that caused it — with shared-span evidence and counterfactual replay proving what would have happened without it.
The written report
A print-ready deliverable your boss can read: executive summary, findings ranked by severity, evidence chains, the honest blind-spot map (what we could NOT verify), and a prioritized fix list.
The readout
A 30-minute call walking your team through every finding and exactly how to fix each one. Re-run the same commands after the fix to prove it stuck.
How it works
Kickoff to fix list in 48 hours.
15-minute fit call
Free. We confirm the integration path for Mem0, Zep, Letta, Cloudflare Agents, or your vector store; AgentCore engagements are assessed case-by-case through the benchmark harness. We agree on scope and sign NDAs if needed. If we're not a fit, we tell you plainly.
Kickoff + runs
We instrument together on a call, seed canaries, and run the battery in your staging or prod-adjacent environment. Nothing leaves your infrastructure — Ferryte is source-available, so your security team can read every line first.
Report + readout
You get the written report and a 30-minute readout: every finding, its evidence, its fix — and what we couldn't verify, stated honestly.
The deliverable
Read a sample before you buy.
This is a real report generated by the same pipeline your audit uses, run against our reproducible multi-tenant-leak demo and anonymized. Every claim in your report carries its evidence: the shared span, the retrieval trace, the supersession edge, the replay verdict.
The guarantee
Nothing actionable found?
Full refund. No argument.
If the audit surfaces no actionable finding — no structural failure, no attributable wrong answer, no blind spot worth closing — you pay nothing. Audit customers also get first claim on the five design-partner seats: six months of Cloud free when it ships, and the audit fee credited against it.
Two audit slots per week. First reply within 24 hours.
FAQ
What buyers ask first.
Does our data leave our infrastructure?
No. Ferryte runs inside your environment; memory contents never leave it. The only artifact that crosses the boundary is the report itself, and you review it before anyone else sees it. Ferryte is source-available (BSL 1.1), so your security team can audit every line of what runs.
What do you need from us?
A Python agent with a persistent memory backend, one engineer on the kickoff call, access to a staging or prod-adjacent environment, and — ideally — up to three wrong answers you want explained. That's it.
Our memory layer is custom / not on your list.
If it's a vector store with write, search, and delete, the generic adapter usually covers it — that's exactly what the free fit call establishes. If we can't support your stack, we say so on that call and you pay nothing.
Why is it so cheap?
Because it's also how we choose design partners for Ferryte Cloud. You get a senior engineer and the full engine at a price that needs no procurement; we get to see real production memory bugs. The fee is credited against Cloud if you take a design-partner seat.
Can you also fix what you find?
The audit hands you the fix list and proves each fix when you apply it. If you want us to implement the fixes with you, that's the Fix Sprint — $1,500, scoped on the readout call.