Portable Agent Memory Protocol Enables Cryptographically-Verified AI Agent Memory Transfer
Researchers have introduced a novel open protocol named Portable Agent Memory (PAM), which facilitates the transfer of persistent memory states among diverse AI agents. Published on arXiv, PAM tackles the issue of AI agent context—including episodic events, semantic knowledge, procedural skills, working state, and identity preferences—being confined within specific vendor environments. The protocol incorporates a structured memory model with five components, utilizing a Merkle-DAG provenance graph for tamper evidence, capability-based access control for selective memory sharing, and an injection-resistant rehydration method that adjusts recalled content for various target models while reducing the risk of indirect prompt injection. Additionally, it features a JSON-first serialization format with optional CBOR compaction. A Python SDK, complete with 54 successful tests, and an agent ski (possibly a typo for 'skill' or 'kit') are also available. PAM aims to enhance interoperability and liberate AI agent memory from proprietary limitations.
Key facts
- Portable Agent Memory (PAM) is an open protocol for transferring memory across heterogeneous AI agents.
- Modern AI agents accumulate context (episodic events, semantic knowledge, procedural skills, working state, identity preferences) locked in vendor-specific runtimes.
- PAM uses a five-component structured memory model with content-addressable entries linked by a Merkle-DAG provenance graph.
- The protocol includes capability-based access control for selective memory disclosure.
- An injection-resistant rehydration protocol adapts recalled content to heterogeneous target models.
- PAM supports JSON-first serialization with optional CBOR compaction.
- A Python SDK with 54 passing tests is available.
- The protocol aims to enable cryptographically-verified memory transfer across different AI agent systems.
Entities
—