Durmem is tested on three key AI memory benchmarks: one for consistency as facts change, one for recall across long histories, and one for memory over real agent work. Together they ask whether an agent can recover the right current context when it matters.
Changing-history memory. Same-cohort reference run: CUPMem at 78.7%.
Long-horizon memory QA. Published full-context comparator: GPT-4o at 60.6%.
Context: Some vendors publish results in the mid-to-high 90s, often after heavy benchmark-specific optimization. Durmem did not engage in benchmark-specific tuning; this score reflects its general memory system.
Memory over realistic agent work. Published full-context comparator: GPT-5.2 at 72.3%.
The table keeps the homepage figures in one place and shows reference points from the same disclosed cohort or the corresponding benchmark papers. It is not a cross-suite ranking: cohorts and model backbones may differ.
| Benchmark | Durmem cohort result | Reference points | What the benchmark stresses |
|---|---|---|---|
| STALE | 86.1% | Same 141-scenario cohort: CUPMem 78.7%, GPT-4o-mini without memory 8.7%, Mem0 8.3%, LiCoMemory 7.6%, Zep 6.0%. | Whether memory stays current as facts change, including stale premises and cascading conflicts. |
| LongMemEval | 89.3% | Published full-context reference: GPT-4o 60.6%. | Long-horizon questions over human history, including updates and temporal reasoning. |
| AMA-Bench | 75.6% | Published references on their reported evaluation: GPT-5.2 full-context 72.3%, AMA-Agent 57.2%, MemoRAG 46.1%. | Whether agents can use memory from prior tool use, task steps, preferences, and outcomes. |
Durmem figures are from disclosed evaluation cohorts: 141 of 400 STALE scenarios, 149 LongMemEval questions, and 43 of 208 AMA-Bench cases. STALE references above use Durmem's same 141-scenario cohort; other reference numbers are published results, not systems re-run by Durmem. Cohorts and model backbones may differ, so these are context rather than a whole-field ranking.
A single retrieval score cannot tell the whole story. These benchmarks cover three failure modes an agent encounters in practice: believing something obsolete, losing facts in a long history, and forgetting what happened while another agent was working.
Each scenario contains an earlier fact and a later update that makes it obsolete. The benchmark checks whether the agent identifies the current state, rejects questions built on stale premises, and adapts its response to the new reality.
Questions draw on long, multi-session human histories. Success requires more than finding a sentence: the agent must connect evidence across sessions, order events in time, and honor later knowledge updates.
Instead of human conversation, AMA-Bench uses agent trajectories: actions, tool outputs, observations, state changes, and outcomes. It measures recall, causal inference, state updating, and state abstraction across six domains.
Each evaluation uses a disclosed stratified cohort rather than claiming a full-suite run. The cohorts preserve the relevant conflict types, question categories, agent domains, and memory skills, but sampling uncertainty still applies.
Durmem answers from the structured memory it built during ingestion; it does not receive the original conversation or trajectory at question time. STALE reference scores are Durmem's same-cohort reruns; LongMemEval and AMA-Bench references are published paper figures, not systems re-run by Durmem.
Connect a hosted store, ask your agent to describe it, then use Durmem in the coding workflow you already have open.
Request access Open install guide