What happens when an agent can find the decisions, facts, and instructions a project has already established? It can spend less time searching, repeating work, and recovering from mistakes.
We studied actual Codex interactive sessions, Codex automated sessions, and Claude automated sessions. We then ran controlled comparisons without Durmem’s saved context to estimate the work those same sessions would have needed. The ranges below show conservative and realistic estimates.
Across the observed sessions, the work used 82.1 million tokens with Durmem, including the 7.2 million tokens Durmem itself used. The estimated equivalent without Durmem was 194.7 million to 468.4 million tokens. These results describe the sessions in this study.
The savings, by session
Each row shows actual total tokens used with Durmem and the estimated total for the same work without it. The final two columns turn those estimates into a reduction.
Reduction in total token use
Swipe to see all columns on a small screen.
| Session type | With Durmem | Without, conservative | Without, realistic | Reduction, conservative | Reduction, realistic |
|---|---|---|---|---|---|
| Codex interactive | 27,675,218 | 77,886,608 | 199,011,684 | 64.5% | 86.1% |
| Codex automated | 33,018,064 | 73,339,623 | 172,556,981 | 55.0% | 80.9% |
| Claude automated | 21,430,394 | 43,489,649 | 96,874,051 | 50.7% | 77.9% |
| Combined | 82,123,676 | 194,715,880 | 468,442,716 | 57.8% | 82.5% |
What each Durmem token made possible
The study also compared Durmem’s own token use with the work it helped avoid. Across all three session types, each token Durmem used corresponded to an estimated 16.6 to 54.7 tokens of avoided work.
Tokens saved for each token Durmem used
| Session type | Durmem tokens | Saved, conservative | Saved, realistic | Per token, conservative | Per token, realistic |
|---|---|---|---|---|---|
| Codex interactive | 3,526,219 | 53,737,609 | 174,862,685 | 15.2× | 49.6× |
| Codex automated | 2,309,852 | 42,631,411 | 141,848,769 | 18.5× | 61.4× |
| Claude automated | 1,362,743 | 23,421,998 | 76,806,400 | 17.2× | 56.4× |
| Combined | 7,198,814 | 119,791,018 | 393,517,854 | 16.6× | 54.7× |
Where the savings came from
Shared context helped in five ways. These categories divide the same estimated savings shown above; they are not additional savings. Durmem Kanban also automates the production release, acceptance, and monitoring workflow, with the operator making decisions and giving approvals where required.
Reusing facts and instructions
Already-established knowledge reduces repeated code and history searches, experiments, and questions.
Preventing unsafe or unapproved actions
Decisions and limits reach agents before they act, helping them avoid work that was rejected, postponed, or not approved.
Avoiding duplicate or outdated work
Agents see what was already completed, disproven, or replaced instead of starting over from old assumptions.
Preventing release and multi-computer mistakes
Shared release requirements and computer-specific test evidence help agents avoid incomplete releases, repeated diagnosis, and platform failures.
Avoiding invalid measurements and reruns
Shared study definitions, procedures, and known issues help agents avoid repeating invalid analysis.
How we measured it
Start with actual use
Examine real interactive and automated Codex sessions and automated Claude sessions.
Replay representative cases
Run controlled comparisons without saved context and observe searches, unnecessary work, and questions to the user.
Apply results to similar cases
Count related lookups supporting one decision once, so the same benefit is not credited multiple times.
The conservative estimate values avoided work cautiously and limits credit for each decision. The realistic estimate also accounts for the time serious release, implementation, and measurement mistakes can consume. The figures are estimates for the observed sessions, not a company-wide projection.
Give your agents a shared memory.
Help your team pick up where it left off and keep moving together.
Explore Durmem →