Development evidence
Experiments, including the failures.
What ran, what failed, and what is still unknown. These are not admitted leaderboard results.
UPDATED 30 OCTOBER 2026 · MNEMOSYNE RESULTS
LongMemEval-S: latest recorded results
Three complete 500-question runs from the Mnemosyne repository. Totals rechecked against every recorded verdict.
Same workload, different transport
31 files · hash-bound
Persistent MCP stdio: 320/320 correct firings. CLI: 200/320, with 120 missed short windows. Both recovered every exact-time intention; no duplicate or cancelled firings. Programmed triggers only: this does not measure natural-language memory quality or competitor performance.
Clocked trigger pressure
18 files · hash-bound
Five seeds, 320 live intentions: 207 fired correctly and 113 short-window triggers were missed. All 160 exact-time intentions recovered; no duplicate or cancelled firings. This measures the public CLI path, not intrinsic engine capacity.
Dependency formation
21 files · hash-bound
1 of 100 cases completed. The next case failed on task-ID reuse; 98 were not attempted. The completed case missed one expected firing. No full-corpus score.
JSON versus schema
78 files · hash-bound
22 first-turn responses across 11 inputs. All 11 plain-JSON outputs failed format checks; all 11 schema outputs passed them. Semantic failures remain. No tasks were executed.
Prompt semantics
77 files · hash-bound
22 first-turn responses across 11 inputs. Added instructions corrected one recurrence proposal, but event and condition errors remained and an unrelated task was introduced. No overall improvement established.