AuthorsLauri Lovén, Jaakko Sauvola, Jukka Riekki, Sasu Tarkoma
TL;DR
MELD uses a five-outcome merge-on-receive protocol with Patch and a status CRDT to match centralized recall while beating naive union at 11% less live storage.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Autonomous agents can call each other’s tools over MCP, but no protocol lets their memories reconcile paraphrases, link related facts, or preserve contradictions.
As a result, each wiki brain stays siloed, knowledge is redundantly re-derived, and ad-hoc last writer wins overwrites silently discard information and provenance.
HOW IT WORKS
MELD’s core mechanism is a merge decision procedure that drives an authenticated Patch over a shared status CRDT on a semantic publish subscribe transport.
You can think of MELD as a distributed wiki brain where each agent’s graph is like a card catalog, and the status CRDT is the ledger that keeps all catalogs aligned.
This merge-on-receive loop lets MELD reconcile paraphrases, relate overlapping claims, and preserve conflicts in graph structure, instead of flattening everything into a single context window.
DIAGRAM
This diagram shows how MELD’s merge-on-receive loop classifies an incoming claim into insert, merge, relate, conflict, or reject.
DIAGRAM
This diagram shows how MELD is evaluated across the 5G edge, national HPC, and local tiers using HotpotQA distractor.
PROCESS
Sense Signals
MELD reads claim key identity, embedding similarity, and NLI verdict for each incoming claim against the receiving wiki brain.
Merge Decision Procedure
Using the merge decision procedure, MELD classifies the claim as insert, merge, relate, conflict, or reject under context, authority, and freshness gates.
Emit Patch
MELD reifies the decision as an authenticated Patch that records the outcome, target, emitted deltas, and which gates fired.
Status CRDT Convergence
Status links from Patches are merged via a grow only status CRDT, reconverging per claim status across sovereign brains after partitions and lossy routing.
KEY CONTRIBUTIONS
Wire Protocol For Knowledge Merge
MELD defines a single PUBLISH operation for graph deltas of concepts, claims, and typed links, with status changes carried as append only links and Patch as the only state mutating object.
Operational Merge Decision Procedure
MELD’s merge decision procedure uses claim key identity, embedding similarity, and NLI to separate merge, relate, conflict, insert, and reject, achieving AUC 0.968 with a 0.013 false merge rate.
Status CRDT With Authority Gating
MELD binds merge semantics onto a publish subscribe transport with a per claim status CRDT that reconverges in 30 of 30 partition heal trials, while last writer wins manages 11 of 30.
RESULTS
Recall@5
0.630
+0.011 over centralized store (0.619)
Recall vs Union
0.630
+0.035 over naive union (0.595)
False Merge Rate
0.013
merge classifier on adjudicated candidate pairs
Partition Heal Trials
30/30
status CRDT reconverges vs 11/30 for last writer wins
These numbers come from HotpotQA distractor and a constructed claim pair gold set, testing recall, merge safety, and status convergence. MAIN_RESULT shows that MELD can decentralize agent memories while matching centralized recall and improving over naive union at lower live storage.
BENCHMARK
These numbers come from HotpotQA distractor and a constructed claim pair gold set, testing recall, merge safety, and status convergence. MAIN_RESULT shows that MELD can decentralize agent memories while matching centralized recall and improving over naive union at lower live storage.
BENCHMARK
Recall@5 of gold supporting facts for MELD, centralized store, and naive union.
KEY INSIGHT
MELD’s decentralized merge achieves recall@5 = 0.630, slightly higher than the centralized store’s 0.619, while using fewer live units than naive union.
This is surprising because we usually expect distribution and semantic routing to hurt recall, yet MELD’s merge procedure and status CRDT preserve or improve it.
WHY IT MATTERS
MELD unlocks a practical way for many sovereign agent wiki brains to share and reconcile knowledge without a central coordinator or destructive overwrites.
Builders can now deploy multi agent systems where memories merge, relate, and preserve conflicts across a 5G edge to HPC continuum, instead of staying siloed or relying on brittle last writer wins.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…