TechPrufer
Get started
Milvus
Milvus logoMilvus · Feature trail

Robot Long-Term Memory

1 post · Jun 2023 – Apr 2026

What this trail covers

Milvus powers the long-term memory tier for robot task execution systems, storing semantic information as vector embeddings for retrieval via vector similarity search, often combined with metadata filtering (hybrid search). This enables robots to adapt to environmental changes and improve future task performance. The integration with LangChain 1.0 further enhances this by providing a production-ready framework for building agents with persistent, intelligent memory, leveraging Milvus for efficient storage and retrieval. This post details how OpenAgents, a framework for distributed multi-agent collaboration, integrates with Milvus to provide a scalable, high-performance long-term memory layer for AI agent networks. Milvus enables fast semantic search, billion-scale horizontal scalability, and multi-tenant isolation through Partition Key, allowing agents to store, retrieve, and reuse knowledge effectively. The integration is achieved through custom Mods that call Milvus APIs directly, storing embedded agent messages, tool outputs, and interaction logs. Developers can customize embedding models, storage schemas, and retrieval strategies. A demo showcases a developer-support community where specialist agents collaborate using Milvus for memory, remembering past answers, reusing explanations, and improving over time.

All Milvus trails
  1. How RoboBrain Builds Long-Term Robot Memory with MilvusThis post details how RoboBrain, an embodied-intelligence operating system for robot task execution, utilizes Milvus as its long-term semantic memory tier. It explains how RoboBrain addresses production failures in robot task execution by implementing task objectification, tiered memory, and a feedback loop. Specifically, it highlights Milvus's role in storing and retrieving scene knowledge, historical experience, exception cases, and post-task writebacks as vector embeddings, enabling semantic similarity search and hybrid search (combining vector similarity with metadata filtering) for robots to adapt to environmental changes and improve future task performance. The post contrasts Milvus's role with relational databases for task IDs and logging systems for raw logs, emphasizing Milvus's suitability for recalling information by meaning rather than by structured fields.
  2. How to Add Long-Term Memory to Anthropic's Managed Agents with MilvusThis post details how Milvus can be integrated with Anthropic's Managed Agents to provide two key memory capabilities: 1) Intra-session semantic recall by indexing high-signal events (decisions, strategies, resolutions, error handling) into a Milvus collection alongside the session log, enabling agents to query their own past decisions. 2) Inter-session shared experience by creating a partitioned Milvus collection (partitioned by tenant_id) where agent experiences are stored upon session completion and retrieved at startup for new sessions, allowing parallel agents to learn from each other. It highlights production considerations such as selecting indexable events, defining consistency boundaries, and controlling embedding costs for intra-session memory, and the use of Milvus partition keys for tenant isolation in shared memory.
  3. Fix RAG Retrieval Errors with CRAG, LangGraph, and MilvusThis post details the implementation of a Corrective Retrieval-Augmented Generation (CRAG) system using LangChain, LangGraph, and Milvus. It addresses common RAG failures like retrieval mismatch, stale content, and memory contamination by introducing an evaluation and correction step. The implementation leverages Milvus's Partition Key for multi-tenant isolation, hybrid retrieval (dense vectors, sparse vectors for BM25, and scalar metadata filtering) with Reciprocal Rank Fusion (RRF) for result fusion, and dynamic JSON fields for flexible schema evolution. The post contrasts a middleware approach with a state-graph approach in LangGraph for building CRAG systems, favoring middleware for its reduced code, easier debugging, and centralized control flow. It provides environment setup instructions and a Python code example demonstrating the CRAG middleware integration with Milvus for retrieval and Tavily Search for web fallback.
  4. How to Build Production-Ready AI Agents with Deep Agents and MilvusThis post details how Deep Agents, a LangChain framework, can be combined with Milvus to build AI agents with structured workflows and long-term memory. It explains Deep Agents' capabilities (task planning, file system access, sub-agent delegation) and customization options (system prompt, tools, middleware, sub-agents, interrupt_on, backend). It highlights Milvus's suitability for persistent memory due to its compute-storage separation, horizontal scaling, high-concurrency queries, and real-time updates. The integration uses a CompositeBackend to route temporary data to StateBackend and persistent memory data to StoreBackend backed by Milvus, enabling agents to store and retrieve embeddings for cross-session memory.
  5. How to Build Production-Ready AI Agents with Long-Term Memory Using Google ADK and MilvusThis post details the integration of Google ADK with Milvus for building production-ready AI agents with explicit separation of short-term session memory and long-term memory. It demonstrates how Milvus's semantic retrieval and hybrid query capabilities are leveraged for agent memory backends, including filtering by structured fields like user_id. The post also covers ADK's three-layer architecture, event recording for replayability, and prefix-based data scoping for visibility control. It provides a practical example of building a technical support agent using Milvus Lite/Standalone for local development and Docker Compose for Milvus deployment.
  7. We Extracted OpenClaw’s Memory System and Open-Sourced It (memsearch)Introduces memsearch, a standalone, plug-and-play long-term memory library for AI agents. It extracts and open-sources the memory architecture from OpenClaw, which uses Markdown files as the source of truth and Milvus for vector indexing. Key contributions include: 1. Markdown-first memory: All memory stored as plain Markdown files, enabling human readability, editability, and Git versioning. 2. Transparent indexing: Milvus index is not the source of truth; Markdown files are reconstructible. 3. Effortless migration: Plaintext memory allows easy switching of machines, embedding models, or vector databases. 4. Human-AI co-authorship: AI handles daily logs, humans manage long-term principles in shared Markdown files. 5. Core workflows: Watch (1500ms debounce for file monitoring), Index (semantic chunking, SHA-256 deduplication, versioned chunk IDs including model version), Search (retrieval), and Compact (summarization).
  8. Why Clawdbot Went Viral — And How to Build Production-Ready Long-Running Agents with LangGraph and MilvusThis post details how Milvus, specifically Milvus Lite and distributed Milvus, can be used to provide persistent semantic memory for long-running AI agents. It explains the memory problem faced by agents (context exhaustion, forgetting across sessions) and how vector databases like Milvus solve this by storing and retrieving semantic embeddings of progress records and git commits. The post also introduces a two-agent architecture (initializer and coding agent) for managing complex tasks and emphasizes browser automation for verification. It highlights Milvus's suitability for production-scale vector search, its hybrid search capabilities, and the ease of migration from Milvus Lite to distributed Milvus.
  9. Building AI Agents in 10 Minutes Using Natural Language with LangSmith Agent Builder + MilvusThis post demonstrates the integration of LangSmith Agent Builder with Milvus to create AI agents with persistent memory and tool-use capabilities. It highlights how natural language descriptions can be used to define agent goals and tools, with Milvus serving as the vector database for memory. The post details the process of setting up API keys, creating an agent, and configuring tools, including the use of Milvus's MCP server for vector search. It also contrasts Agent Builder with DeepAgents, positioning Agent Builder for rapid prototyping and ease of use, while DeepAgents offers more granular control for complex, production-grade agents. The core contribution is showcasing a no-code approach to building memory-enabled AI agents that leverage Milvus for context recall and preference persistence.
  10. LangChain 1.0 and Milvus: How to Build Production-Ready Agents with Real Long-Term MemoryThis post details how LangChain 1.0's new ReAct agent architecture, middleware, and structured output capabilities, when combined with Milvus as a long-term memory store, enable production-ready AI agents. It highlights the shift from flexible Chains to a standardized ReAct loop, the introduction of Middleware for production-grade control (PII detection, summarization, retries), and standardized structured output via Pydantic models. The integration with Milvus is presented as a key enabler for persistent, intelligent memory in agents, allowing them to store, recall, and reuse information efficiently via semantic similarity search.
  11. OpenAgents x Milvus: How to Build Smarter Multi-Agent Systems That Share MemoryThis post details the integration of OpenAgents, a framework for distributed multi-agent collaboration, with Milvus to provide a scalable, high-performance long-term memory layer for AI agent networks. It explains how Milvus's semantic search, billion-scale scalability, and multi-tenant isolation (via Partition Key) are leveraged. The integration is implemented via custom Mods that embed agent messages, tool outputs, and interaction logs into Milvus for storage and retrieval. The post also provides a practical example of building a developer-support community using this integration, highlighting how Milvus enables agents to remember past interactions and improve over time. It outlines the necessary dependencies (openagents, pymilvus, sentence-transformers, openai, python-dotenv) and environment variable configurations for Milvus URI, OpenAI API keys, and embedding models. The configuration also details the OpenAgents network setup, including transport protocols (HTTP) and Mods for workspace messaging (general, coordination, expert channels).
  12. Conversational Memory in LangChainThis post details how to integrate LangChain's conversational memory with Milvus Lite. It demonstrates setting up `OpenAIEmbeddings`, `OpenAI` LLM, `VectorStoreRetrieverMemory`, and `ConversationChain` to store and retrieve conversational history. The example shows saving context (favorite snack, sport, beer, dessert, musician) and then querying for this information, confirming Milvus's role in providing persistent memory for LLM applications.