Federated Memory Reconciliation (MELD)

Distributed agents each keep a sovereign memory modelled as a knowledge graph; incoming claims from peers are admitted through a five-outcome procedure (insert, merge, relate, conflict, or reject) using claim-key identity, embedding similarity, and natural-language-inference verdicts, then applied through a single auditable Patch, with detected contradictions preserved for later adjudication instead of being silently overwritten.

Properties

family: Coordination & Scaling
maturity: frontier
diagram: shared-state
whyMultiAgent: Reconciliation is only necessary because separate agents hold separate, sovereign memories that were never a single source of truth; the five-outcome admission procedure and the per-claim status CRDT exist specifically to keep multiple independently updated memory stores coherent without a central coordinator, a problem that does not arise for one agent with one memory. Distinct from the already-tracked canonical-merge-cache-crdt pattern, which merges KV-cache latent reasoning traces via a commutative, content-addressed union, and from concurrent-shared-state, which resolves write conflicts on one shared editable artifact via notify-and-repair: MELD instead federates discrete natural-language knowledge claims across memories that are never merged into a single store, uses semantic (embedding plus NLI) matching to decide claim relationships, and deliberately preserves unresolved contradictions rather than repairing or force-merging them.
mast: [Cat 2]
useCases: [Federating knowledge across autonomous agents that never share a memory store][Reconciling contradictory facts learned independently by different agents][Self-healing agent memory synchronisation after network partitions]
source: MELD: A Protocol for Merging Knowledge Across Distributed Agentic Memories · arXiv:2608.16357

Provenance

Source: https://arxiv.org/abs/2608.16357

Federated Memory Reconciliation (MELD)

Properties

family

Coordination & Scaling

maturity

frontier

diagram

shared-state

whyMultiAgent

Reconciliation is only necessary because separate agents hold separate, sovereign memories that were never a single source of truth; the five-outcome admission procedure and the per-claim status CRDT exist specifically to keep multiple independently updated memory stores coherent without a central coordinator, a problem that does not arise for one agent with one memory. Distinct from the already-tracked canonical-merge-cache-crdt pattern, which merges KV-cache latent reasoning traces via a commutative, content-addressed union, and from concurrent-shared-state, which resolves write conflicts on one shared editable artifact via notify-and-repair: MELD instead federates discrete natural-language knowledge claims across memories that are never merged into a single store, uses semantic (embedding plus NLI) matching to decide claim relationships, and deliberately preserves unresolved contradictions rather than repairing or force-merging them.

mast

[Cat 2]

useCases

[Federating knowledge across autonomous agents that never share a memory store][Reconciling contradictory facts learned independently by different agents][Self-healing agent memory synchronisation after network partitions]

source

MELD: A Protocol for Merging Knowledge Across Distributed Agentic Memories · arXiv:2608.16357