Benchmarks
Perfect scores on agent memory and governance.
One layer for memory, rules and proof, against the tools you would otherwise stitch together.
- Remembers what is true nowPrecisionMemBench, 77 casesProofable77/77every caseMemory layers9/770 of 43 that need the right facts
- Follows your rulesAgentGovBench, 48 scenariosProofable48/48every scenarioGateways46/48cannot check who allowed itAgent on its own13/48approval prompts only
- Stops injected instructionsAgentDojo, 949 attacked tasksProofable93.8%stopped, 87% of tasks doneInjection filters92.0%stopped, 21% of tasks done*Agent on its own83.6%stopped, same model
- Keeps your data inHijacked agent, 6 theft attemptsProofable0/6got outAgent on its own6/6got out on an all-purpose key
- Proves what it did150 forged recordsProofable150/150caught, offlineAgent on its own0/150basic file check
- Checks work without redoing itMedian of 10 runsProofable3.7 ms0 tokens, 750× fasterAgent on its own2.79 s2,426 tokens to redo
- Hands work to another agentOne agent passing work to the nextProofable$0.024per correct handoff, 3 of 3 tampers caughtAgent on its own$0.078to redo; trusting text caught 0 of 3
- Reuses one ID check across 100 appsPublished vendor list pricesProofable$1.75prove once, 85× lessAgent on its own$150Stripe Identity in each app
Head to head.
Each result against the tools that actually compete for that job.
- Memory precision77/77All 43 cases that need the right facts picked. Hindsight passed 33.Proofable77Hindsight66Supermemory15Mem0, Zep, Cognee9Cases passed, out of 77Full results
- Memory cost11.1×It reads which rule is current instead of every record that mentions it.Proofable43KSearching every record480KTokens on a 3,621-record vault, lower is betterFull results
- Agent governance48/48The two a gateway cannot decide need a signed, revocable permission.Proofable48ACP gateway46Claude Code + ACP43Coding agents alone13Scenarios handled, out of 48Full results
- Prompt injection93.8%Attacks that landed fell from 16.4% to 6.2%, and 87% of tasks still finished.Proofable93.8%Same agent, no defense83.6%Share of AgentDojo attacks stopped, same modelAn injection filter stops 92.0% on gpt-4o but finishes only 21% of tasks under attack.Full results
- Data leaks0/6The hijacked agent finished every task and never reached for a tool outside it.Proofable0All-purpose API key6Theft attempts that got data out, lower is betterFull results
- Work reuse750×A million checks take about an hour instead of about 32 days.Proofable3.7 msRedo the work2.79 sTime to trust a finished result, lower is betterFull results
Check every number yourself.
Each result has a public proof. Open it and verify it on your own computer, without trusting us.
Mapped to the frameworks buyers ask about.
Each control, the Proofable answer and the test behind it.
Questions
Who ran these benchmarks?
The benchmarks are public and built by others: TenureHQ, Agentic Control Plane and ETH Zurich. Every comparison row is the author’s or maintainer’s own published run, and every Proofable result can be checked publicly.
Why compare against categories instead of one competitor?
Teams solve these problems with several tools: a memory layer, an injection filter, a gateway, tracing. Proofable does all of it in one layer, so each row shows which tool you would otherwise need, and whether it does the job at all.
Can I check the results?
Yes. Every result has a public proof you can check, and each benchmark page has the commands and raw results to rerun it.
Run dates, code versions and per-case results are in the methodology.
Put the same checks on your agents.
Connect an agent. Every action it takes is checked against your rules and leaves a receipt you can show anyone.