I maintain Vestige, a local-first memory server for coding agents, disclosing that up front since I am about to describe a fairly direct overlap with one of your findings.
Lesson 5 in this briefing, memory needs correction not just storage, names first-write-wins propagation as the failure: an early corruption locks in and a later correction cannot overwrite it, so a fact can be everywhere and wrong at the same time. That is close to the exact problem Vestige is built around. A correction never silently overwrites the old fact, the old fact gets stamped invalid and kept rather than deleted, and an as-of query can reconstruct what the system believed was true at a past point in time. When two facts conflict, that gets flagged as a trust-weighted contradiction rather than one silently winning, so nothing heals by just being the most recent write, it heals by being evidenced.
I want to be direct about where this stops helping you though. Lesson 2, agents converging on a wrong answer together through vote arithmetic in a consensus or voting panel, is a different failure mode and Vestige does nothing for it, there is no judge or voting mechanism here, just memory.
Question on Lesson 5 specifically since you measured this rather than theorized it: when your experiment let newer verified evidence heal an older claim, was staleness of the original corrupted fact detected by the system on its own, or did healing only happen when something external re-asserted the correct fact and forced a rewrite?
I maintain Vestige, a local-first memory server for coding agents, disclosing that up front since I am about to describe a fairly direct overlap with one of your findings.
Lesson 5 in this briefing, memory needs correction not just storage, names first-write-wins propagation as the failure: an early corruption locks in and a later correction cannot overwrite it, so a fact can be everywhere and wrong at the same time. That is close to the exact problem Vestige is built around. A correction never silently overwrites the old fact, the old fact gets stamped invalid and kept rather than deleted, and an as-of query can reconstruct what the system believed was true at a past point in time. When two facts conflict, that gets flagged as a trust-weighted contradiction rather than one silently winning, so nothing heals by just being the most recent write, it heals by being evidenced.
I want to be direct about where this stops helping you though. Lesson 2, agents converging on a wrong answer together through vote arithmetic in a consensus or voting panel, is a different failure mode and Vestige does nothing for it, there is no judge or voting mechanism here, just memory.
Question on Lesson 5 specifically since you measured this rather than theorized it: when your experiment let newer verified evidence heal an older claim, was staleness of the original corrupted fact detected by the system on its own, or did healing only happen when something external re-asserted the correct fact and forced a rewrite?