Self-correction is sold as the cure for agent failure. In practice, it is the first failure mode to reach production. When one agent owns its own corrections, the loop is trivially bounded. When N agents share memory and correct each other's writes, the loop becomes a spiral — and it can diverge before anyone reads a log line.

The degenerate loop

Three agents, one MCP memory server. Agent C reads a fact from shared memory at T1. At T2, Agent A discovers the fact is stale and writes a correction. At T3, Agent C — still holding the T1 snapshot — finishes its chain, sees the corrected fact conflicts with its snapshot, and "repairs" memory back to the original. Both writes are internally rational. Both are wrong at the system level.

You can observe this. Add a correction-reversal counter to the memory server's audit trail: scan successive writes to the same slot, and if the semantic direction flips (A→B→A), increment. A rising reversal count per slot is the earliest measurement of a degenerate spiral. It is not a theory; it is a delta computation on writes you already log.

Mechanism 1: diff invariants on the write path

Place a gate between agents and persistent memory, implemented as an MCP middleware layer. Every write carries provenance: agent ID, a hash of the reasoning trace that produced the correction, and the agent's declared justification. The gate computes a diff against the last K versions. If the diff reverses a prior correction on the same statement, the write is not applied. It enters a quorum queue.

Quorum here is not democracy. A second, independent agent must reproduce the same correction from different evidence — different tool calls, different inputs — within the timeout window. If only the original agent insists, the write is rejected and the slot is locked until a human inspects it. The rule is blunt but enforceable: shared memory must not oscillate, and anyone who wants to flip a shared fact pays the cost of independent verification.

Mechanism 2: bounded recursion with a frozen reference

Unbounded self-correction has no attractor. Loops converge only if an external reference pins them. Split the MCP memory resource into two layers:

• **

maref #ai #opensource #machinelearning