Feature
Multi-Agent Arbitration
Fleets of long-running agents write to shared memory, and they will disagree — a stale agent re-asserts an old fact, a compromised one injects a wrong one, two honest ones observe different things. Most memory systems resolve this with last-write-wins, which means your shared knowledge is only as reliable as your least reliable agent. MetaMemory arbitrates instead: every write carries verified agent identity, conflicts are weighed by per-agent trust, and an established fact is never silently evicted.
Server-Derived Attribution
Every memory and claim write is attributed to its agent principal from the authenticated token — not from anything the agent says about itself. An agent cannot impersonate another or launder a fact through a different identity.
The Trust Ladder
Each principal carries a trust weight from 0 to 1. Below the supersede floor (0.5), a challenger can never evict another agent’s current fact. At strong trust (0.75+), its claims stand without corroboration. In between, corroboration from independent sources tips the decision — and a contested claim never silently voids the established current value.
Quarantine & Circuit Breakers
A misbehaving agent can be quarantined in one call — its influence stops without deleting its history, so you can investigate what it wrote. Circuit breakers trip on anomalous claim churn, containing a runaway or compromised writer before it corrupts shared knowledge.
Audited Decisions
Every arbitration decision snapshots its inputs at decision time: the trust weights involved, the corroborating sources, and the thresholds applied. When someone asks why the system believes what it believes, the answer is in the ledger, not in a log that rotated away.
Server-derived
Write Attribution
Per-agent, 0–1
Trust Weights
Never void current
Contested Claims
Inputs snapshotted
Decision Audit