AMP: A Vendor-Neutral Wire Format for Agent Memory Operations

Researchers introduce memorywire, a vendor-neutral JSON wire format standardizing agent memory operations across competing frameworks like mem0, MemGPT, and Cognee. The protocol enables interoperability between memory systems while including human-in-the-loop governance controls, with a reference implementation achieving 100% recall on test queries and 68/80 conformance across adapters.

The fragmentation of agent-memory frameworks represents a critical infrastructure gap in the emerging AI agents ecosystem. Each platform—mem0, Letta/MemGPT, Cognee, Zep, MemoryOS, MemTensor—operates in isolation with proprietary SDKs and storage formats, forcing developers into vendor lock-in and making data migration costly and lossy. Memorywire addresses this by proposing a standardized wire format for five core operations (remember, recall, forget, merge, expire) across four memory types (semantic, episodic, procedural, emotional), mirroring cognitive science frameworks rather than engineering convenience.

The technical contribution combines established components—Reciprocal Rank Fusion, finite state machines, and short-term/long-term memory consolidation—into a protocol-agnostic packaging positioned to interoperate with the Model Context Protocol rather than compete with it. Empirical validation shows strong performance: 100% recall@5 on labeled queries with sub-50ms latency on both ingestion and retrieval. The cross-adapter conformance suite (68/80 passing cells) demonstrates real-world viability across heterogeneous backends.

The governance layer—an optional human-in-the-loop approval channel—addresses a critical gap in existing frameworks: no current system provides human oversight before memories enter long-term storage, creating audit and safety risks in production deployments. For developers, this standardization could accelerate agent adoption by reducing integration complexity and enabling memory portability. For enterprises, it reduces switching costs and mitigates vendor dependency. The positioning as complementary infrastructure rather than a standalone product suggests pragmatic design philosophy focused on ecosystem composition rather than market dominance.

• →Memorywire enables memory interoperability across six fragmented agent-memory frameworks through a JSON Schema wire format, reducing vendor lock-in.
• →Human-in-the-loop governance controls address a safety gap—no existing framework provides oversight before memories enter long-term storage.
• →Empirical benchmarks show 100% recall@5 on test queries with sub-50ms latency, validating production-readiness across five backend adapters.
• →The protocol is designed to compose with Model Context Protocol rather than compete, positioning it as ecosystem infrastructure rather than standalone product.
• →Cross-adapter conformance testing (68/80 cells) demonstrates real-world viability despite architectural heterogeneity across participating frameworks.