Unleash the full potential of your AI agents.
Durmem is a shared memory layer that works across harnesses and computers to solve three common pain points for coding agents.
Important memory gets lost after compaction and isn’t shared across harnesses or computers. Developers have to repeat the same instructions and explanations.
Tell it once. Every agent remembers.Durmem automatically captures important decisions and instructions in durable memory shared across computers and harnesses, so connected agents with access to that memory can recall what you’ve already explained.
As context fades during long-running, multi-agent work, agents lose sight of the goal, drift from key decisions, and go off track.
Keep agents focused as they work.Durmem automatically brings relevant decisions, instructions, and direction back into context as agents work—helping them keep complex tasks on course through context resets and handoffs.
Agents make conflicting decisions, rediscover known facts, and contradict earlier decisions—leading to wasted work or the wrong implementation.
Harmony, not chaos. Significant token savings.Durmem keeps agents aligned with shared facts and decisions, so they build on each other’s work instead of fighting it—cutting rework and wasted tokens.
Up to 80% token savings, estimated from real-world coding workloads.
Each Durmem store preserves a team’s knowledge, working principles, and boundaries. One agent can connect to multiple stores, so it can work across teams with the right context for each one—following local rules without mixing everything into one giant pool.
Compare the same work without Durmem’s shared memory and with the right context already available.
Illustrative workflow. Token counts show the example’s trade-off, not measured benchmark results.
| Service | Monthly cost |
|---|---|
| AWS | $340 |
| OpenAI | $120 |
| Cursor | $60 |
Durmem reports results from disclosed cohorts of three major AI memory benchmarks alongside clearly labeled same-cohort or published-paper reference points. Together, they test whether memory stays current, survives long histories, and supports real agent work.
Tests whether memory recognizes superseded facts and adapts when new information changes what is true. On the same disclosed 141-scenario cohort, the CUPMem reference run scored 78.7%.
Benchmark detailsTests recall and reasoning across long, multi-session histories, including temporal questions and knowledge updates. GPT-4o full-context: 60.6%.
Benchmark detailsTests memory over agent actions, tool outputs, state changes, and outcomes across realistic tasks. For context, the AMA-Bench paper reports GPT-5.2 full-context at 72.3%; cohorts and model backbones differ.
Benchmark detailsDurmem figures reflect 141 of 400 STALE scenarios, 149 LongMemEval questions, and 43 of 208 AMA-Bench cases. See the benchmark details and partial reproducibility receipt. STALE comparison figures are same-cohort reruns; LongMemEval and AMA-Bench references are published paper figures (STALE paper: Chao et al., arXiv:2605.06527 · LongMemEval: Wu et al., arXiv:2410.10813 · AMA-Bench: arXiv:2602.22769, ICML 2026).
Connect Durmem and give your coding agents a shared memory in two steps.
One OS-specific command installs Durmem for your user account. No coding agent and no signed-in account required yet.
durmem setupChoose which detected harnesses to connect—Claude Code, Codex, or OpenCode. Setup signs you in when needed, connects your stores, and installs and verifies the selected integrations. Then start a new harness session.
Start with free hosted Durmem, or choose self-managed Durmem on your own infrastructure. A paid hosted tier with more capacity will be offered later.
A paid hosted tier with more stores and storage will be offered later.
Create your stores, connect your coding agents, and stop rebuilding context from scratch.
Create your stores and get a ready-to-paste connection bundle.