Evolving AI Agent Memory: Introducing Agent Memory…
0 reactions
Evolving AI Agent Memory: Introducing Agent Memory Protocol v1.1
AI agents are only as capable as their context—but managing stateful, long-term memory across complex enterprise deployments has remained a fragmentation bottleneck.
Today, we are excited to share the launch of \*\*Agent Memory Protocol (AMP) v1.1\*\*, representing a major architectural evolution in how persistent cognitive memory is structured, isolated, and scaled.
\#### The Evolution: From MCP Tools to Service-First
In v1.0, AMP was modeled purely as a Model Context Protocol (MCP) toolset. While perfect for local prototyping, this created system-level bottlenecks: low-level DB operations (like consolidation or stats) were exposed directly to the LLM's prompt, bloating context and increasing cognitive load.
\*\*AMP v1.1 solves this by transitioning to a Service-First architecture (HTTP REST / gRPC first) with an optional MCP Tool Adapter.\*\*
This separates concerns cleanly:
1. \*\*Agent-Facing Tools:\*\* Clean cognitive hooks (\`encode\`, \`recall\`, \`forget\`) mapped directly to the LLM context.
2. \*\*Harness-Facing APIs:\*\* Background operations (\`consolidate\`, \`pin\`, \`stats\`) handled out-of-band by the application orchestration framework (LangChain, LlamaIndex, Letta).
\---
\#### Key Enhancements in v1.1
\* \*\*Dual-Delivery Channel Paradigm:\*\* Run the exact same memory contract in two ways. Use the lightweight \*\*MCP Adapter Channel (STDIO/SSE)\*\* for rapid local development, and scale instantly to the \*\*Standalone REST/gRPC API Channel\*\* for production-grade microservices without rewriting a single schema.
\* \*\*Multi-Dimensional Scoping:\*\* Moving beyond single \`agent\_id\` isolation. v1.1 standardizes intersection-based scoping across \`org\_id\`, \`app\_id\`, \`user\_id\`, \`session\_id\`, \`agent\_id\`, \`group\_id\`, and \`workspace\_id\` to natively power collaborative multi-agent workspaces.
\* \*\*Reserved Metadata Vocabulary Registry:\*\* Eliminating database-specific fragmentation. Standardizing properties like TTL, confidence scores, extracted entities, and Subject-Predicate-Object graph relationships (\`amp.relations\`) directly in the schema.
\* \*\*Memory Exchange Format (MXF):\*\* Frictionless NDJSON-based migrations. Back up memory states from platforms like Supermemory or Zep and restore them into local implementations (like \`smriti-memcore\`) with absolute structural fidelity.
\---
\#### 🤝 Built by the Community, For the Community
AMP is an open standard designed to ensure complete backend interoperability. Whether you are building single-user productivity loops or high-throughput enterprise agent platforms, AMP v1.1 provides the robust database-agnostic interface required to manage persistent cognitive state.
Special thanks to Shivam Tyagi, Brad Jones, and the incredible open-source contributors driving this draft forward.
🔗 Explore the full specification and reference implementations on GitHub: https://github.com/smriti-memcore/amp/blob/main/spec/amp-v1.1.md
\#AIAgents #GenerativeAI #SoftwareArchitecture #Microservices #OpenSource #ArtificialIntelligence
#technology
AI agents are only as capable as their context—but managing stateful, long-term memory across complex enterprise deployments has remained a fragmentation bottleneck.
Today, we are excited to share the launch of \*\*Agent Memory Protocol (AMP) v1.1\*\*, representing a major architectural evolution in how persistent cognitive memory is structured, isolated, and scaled.
\#### The Evolution: From MCP Tools to Service-First
In v1.0, AMP was modeled purely as a Model Context Protocol (MCP) toolset. While perfect for local prototyping, this created system-level bottlenecks: low-level DB operations (like consolidation or stats) were exposed directly to the LLM's prompt, bloating context and increasing cognitive load.
\*\*AMP v1.1 solves this by transitioning to a Service-First architecture (HTTP REST / gRPC first) with an optional MCP Tool Adapter.\*\*
This separates concerns cleanly:
1. \*\*Agent-Facing Tools:\*\* Clean cognitive hooks (\`encode\`, \`recall\`, \`forget\`) mapped directly to the LLM context.
2. \*\*Harness-Facing APIs:\*\* Background operations (\`consolidate\`, \`pin\`, \`stats\`) handled out-of-band by the application orchestration framework (LangChain, LlamaIndex, Letta).
\---
\#### Key Enhancements in v1.1
\* \*\*Dual-Delivery Channel Paradigm:\*\* Run the exact same memory contract in two ways. Use the lightweight \*\*MCP Adapter Channel (STDIO/SSE)\*\* for rapid local development, and scale instantly to the \*\*Standalone REST/gRPC API Channel\*\* for production-grade microservices without rewriting a single schema.
\* \*\*Multi-Dimensional Scoping:\*\* Moving beyond single \`agent\_id\` isolation. v1.1 standardizes intersection-based scoping across \`org\_id\`, \`app\_id\`, \`user\_id\`, \`session\_id\`, \`agent\_id\`, \`group\_id\`, and \`workspace\_id\` to natively power collaborative multi-agent workspaces.
\* \*\*Reserved Metadata Vocabulary Registry:\*\* Eliminating database-specific fragmentation. Standardizing properties like TTL, confidence scores, extracted entities, and Subject-Predicate-Object graph relationships (\`amp.relations\`) directly in the schema.
\* \*\*Memory Exchange Format (MXF):\*\* Frictionless NDJSON-based migrations. Back up memory states from platforms like Supermemory or Zep and restore them into local implementations (like \`smriti-memcore\`) with absolute structural fidelity.
\---
\#### 🤝 Built by the Community, For the Community
AMP is an open standard designed to ensure complete backend interoperability. Whether you are building single-user productivity loops or high-throughput enterprise agent platforms, AMP v1.1 provides the robust database-agnostic interface required to manage persistent cognitive state.
Special thanks to Shivam Tyagi, Brad Jones, and the incredible open-source contributors driving this draft forward.
🔗 Explore the full specification and reference implementations on GitHub: https://github.com/smriti-memcore/amp/blob/main/spec/amp-v1.1.md
\#AIAgents #GenerativeAI #SoftwareArchitecture #Microservices #OpenSource #ArtificialIntelligence
#technology
| 3,000 mlx | |
| 5 | |
| 0 |