Title: MemoryWire: A Universal Protocol for Agent Memory Interactions

Original: arXiv:2606.01138v1 Announcement: Cross-domain Abstract: Current agent-memory systems—including mem0, Letta/MemGPT, Cognee, Zep/Graphiti, MemoryOS, and MemTensor—operate in isolation, each employing distinct SDKs, storage architectures, and operational terminologies. The absence of a standardized wire format necessitates custom integrations for every connection, forces complete memory reconstruction during migrations, and lacks a governance mechanism allowing human oversight prior to long-term storage. To address these limitations, we introduce memorywire, a wire format based on JSON-Schema 2020-12 designed to facilitate five core memory operations (remember, recall, forget, merge, expire) across four memory categories (semantic, episodic, procedural, emotional). This solution incorporates a MemoryStore interface, a fan-out router, and an optional Human-in-the-Loop (HITL) governance channel. Our work includes an open-source reference implementation featuring five backend adapters (sqlite-vec, mem0, Letta, Cognee, pgvector). Microbenchmarking on a labeled corpus of 100 facts and 50 queries yielded a perfect recall@5 of 1.000 for 42 queries, with median ingestion and recall times of 37.8 ms and 40.6 ms, respectively. Furthermore, an adversarial-fusion test demonstrated that Reciprocal Rank Fusion maintained a recall@5 of 1.000 throughout a 1-of-N rank-0 injection sweep (K in {0,5,...,50}), whereas max fusion dropped to 0.500 with an 80% error rate when K was 5 or higher. Additionally, a 16-scenario cross-adapter conformance suite successfully passed 68 out of 80 test cases with no failures. Our primary contribution lies not in novel algorithms, but in the standardization of established techniques—such as RRF, finite state machines, STM/LTM consolidation, and diff-and-approve workflows—into a vendor-neutral protocol. Supported by empirical validation, this protocol is designed to complement the Model Context Protocol rather than rival it.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC