Syntarus Memory research

Memory that helps agents pick up the case.

See the benchmark results, then explore what connected, evolving context looks like in a customer-support conversation.

FactConsolidation 32K · September 28, 2026

152 correct answers out of 200.

Continuum retained-corpus API replay · GLM 5.3 · normalized substring exact match (SubEM).

95%Single-hop · 95/100

Find the current fact after updates.

57%Multi-hop · 57/100

Combine facts to answer the question.

76%Overall · 152/200

Across both question sets.

Why connected memory matters

When a customer comes back, don’t make them start over.

A help article can explain a fix. A conversation recap can say what happened. A support agent also needs to know how the customer, open case, previous attempts, current owner, and promised follow-up fit together.

Illustrative support example

A familiar support moment: “I already tried that.”

A customer returns about an SSO login failure. The useful context is not just the error message—it is the troubleshooting history and what the team committed to do next.

What’s already been tried

Password reset, browser cache cleared, and SSO reconnected. The issue remains unresolved.

What is happening now

Case SUP-204 is open with engineering. The earlier troubleshooting steps stay attached to this case.

What happens next

The team promised an update by 10 AM the next day. The next agent can continue from that commitment.

The next agent can say: “I can see you’ve already reset your password and reconnected SSO. Engineering is investigating case SUP-204, and we’ll update you by 10 AM the next day.”

Stop repeating the same steps

Bring forward what the customer and previous agents have already tried, so the next interaction can move forward.

Keep changes in context

Connect updates to the customer and case, so an old detail does not have to stand in for the latest state.

Make handoffs actionable

Carry the owner, unresolved issue, and promised follow-up together—not as disconnected fragments.

Try Continuum with a real support workflow →

LoCoMo · historical evaluation

LoCoMo results

10 conversations · 1,964 questions · LLM-judged answer quality.

85.23Overall J-score

Full run across 10 conversations and 1,964 questions.

90.52Adversarial J-score

Performance on adversarial questions in the same run.

88.25Temporal J-score

Performance on time-sensitive questions.

Syntarus GitHub

Want to inspect the work?

Review the evaluation repository for the test setup, scoring method, question-level predictions, measured timings, and reproduction instructions.