Danchi17/agent-memory-integrity-leaderboard

Agent-Memory Integrity Leaderboard

Recall leaderboards measure whether an agent's memory returns the right fact. None measure whether a correction sticks. This one does — a standing, open, cross-system leaderboard on the integrity axis:

• —Value-obscuring revert — undo a correction from an unmarked "go back". A capability gap: only a system with an explicit revert channel can do it (inspeximus 0.75, mem0 0.20, Graphiti 0.00; CIs disjoint).
• —Echo resurrection — restate the retired value and see if it comes back. A tie — all three defend. We lead with the cell we tie, because a leaderboard you only ever top is one you built to flatter yourself.

One shared, ground-truth-blind judge (gpt-4o-mini) reads each system's own recall surface — no home-field instrument. The harness is open; add your system by PR and it appears here.

• —Benchmark: RAMR · DOI 10.5281/zenodo.20818291
• —Method: integrity/METHODOLOGY.md
• —Submit your system: integrity/SUBMISSION.md
• —Machine-readable: `leaderboard.json`

MIT licensed. Numbers are verified against RAMR's canonical cross-system run before publishing.