TechPrufer
Get started
Weaviate
Weaviate logoWeaviate · Feature trail

Engram: Managed Memory and Context Service

8 posts · Feb 2026 – Oct 2026

What this trail covers

Engram is now generally available as a managed memory and context service for LLM agents. It provides asynchronous, durable pipelines for extracting, reconciling, and persisting structured memory states from raw agent events. Key features include active memory maintenance, fire-and-forget application layer integration with Temporal-grade durability, ready-to-deploy templates for personalization and multi-agent state, built-in scopes for data isolation, and unified retrieval on Weaviate's hybrid search. The Claude Code plugin leverages Engram to provide LLM agents with long-term memory across sessions and codebases, improving consistency and developer productivity by recalling past decisions and code patterns. Memory is saved asynchronously as a side-effect of working, and recalled silently when relevant.

All Weaviate trails
  1. Persistent memory for Claude Code with Weaviate Engram | WeaviateThis post introduces the Engram plugin for Claude Code, which provides LLM agents with persistent memory across sessions and codebases. It works by hooking into Claude Code sessions: before answering, it searches Engram for relevant memories to pass as context, and after answering, it saves the exchange to Engram for future recall. The plugin aims to make LLM agents more like experienced colleagues by remembering past decisions, code patterns, and project context. It emphasizes that memory is best-effort, won't interrupt the user, and is saved asynchronously to avoid slowing down the user. The post also briefly describes Engram as Weaviate's memory server for LLM agents and applications, highlighting its ability to extract, transform, and store memorable parts of conversations.
  2. Agent Memory with Engram: A Practical Guide | WeaviateThis post provides practical guidance on configuring Engram's memory management. It details how topic descriptions act as extraction prompts, influencing what data is remembered and its shape (e.g., atomic facts vs. prose). It explains the impact of 'what to leave out' rules in descriptions for focused memory capture. The post also elaborates on the configuration fields for topics: Name, Description, User Scoped, Property Scopes, and Bounded. It specifically highlights the guarantees provided by 'Bounded' topics (at most one memory per scope) and 'Scope' (partitioning memories by user or other properties), illustrating their effects with examples of memory updates and counts.
  3. Engram is now Generally Available | WeaviateThis post announces the General Availability (GA) of Engram, the managed memory and context service. It elaborates on the production-readiness achieved through hardening pipelines, improving extraction/transformation efficiency, and ensuring deterministic reconciliation. The post highlights the benefits of Engram over common DIY agent memory solutions, emphasizing active reconciliation, durable asynchronous pipelines, and a unified platform for memory and retrieval. It also provides links to documentation, a deep dive, and the Weaviate Cloud console for getting started.
  4. Build a Coding Assistant with Weaviate MCP: RAG over Code & Docs | WeaviateThis post introduces the integration of Weaviate's built-in Model Context Protocol (MCP) server for Retrieval-Augmented Generation (RAG) over codebases and documentation. It details enabling the MCP server, designing schemas for code and documentation collections, implementing syntactic chunking for Python code, and batch ingestion. The post highlights the benefits of MCP for coding assistants, including cost savings, reduced context staleness, and appropriate granularity, by leveraging Weaviate's hybrid search capabilities.
  5. Your LLM Is Only as Good as What It Retrieves | WeaviateThis post provides a deep dive into retrieval quality as a critical factor for LLM reliability in RAG and multi-agent systems. It identifies five specific retrieval failure modes (retrieval drift, context truncation, stale index poisoning, low-relevance top-k retrieval, inter-agent miscommunication) and explains why scaling LLMs does not solve retrieval problems. It proposes four key areas for improving retrieval quality: embedding model selection, chunking architecture, retrieval strategy (including hybrid search, cross-encoder re-ranking, and relevance thresholding), and index maintenance and freshness. Finally, it outlines a measurement framework for evaluating retrieval quality using context precision and context recall.
  7. Engram: Memory by Weaviate | WeaviateThis post details the internal architecture of Engram, Weaviate's managed memory service. It elaborates on the asynchronous pipeline processing, which includes extract, reconcile, and persist steps, built on Temporal for durable execution. It introduces the concept of 'topics' for organizing memories with configurable scopes (project-wide, user-scoped, property-scoped) and bounded constraints, and explains how these are enforced by Weaviate's multi-tenancy. The post also highlights the low-latency API for adding raw data and the use of semantic search for retrieving memories.
  8. Oh Memories, Where'd You Go | WeaviateThis post details the evolution of Engram's integration with LLMs, specifically addressing the challenge of LLMs defaulting to built-in memory (`MEMORY.md`) over external tools like Engram. Key contributions include: 1. Identifying the need for deterministic triggers for memory recall and saving, moving away from agent-driven decisions. 2. Implementing infrastructure-level hooks for session start and per-turn recall, ensuring context is injected before the LLM processes a prompt. 3. Optimizing save performance by making saves asynchronous ('fire and forget') to reduce session overhead. 4. Proposing a more robust pipeline buffer for automatic message ingestion, preventing context loss during premature session termination. 5. Highlighting the distinction between stable facts (`MEMORY.md`) and reasoning chains/contextual evolution (Engram's domain).
  9. Building A Legal RAG App in 36 Hours | WeaviateThis post details the architecture and implementation of a legal RAG application built in 36 hours using Weaviate's Query Agent. It highlights the use of multimodal PDF ingestion with Muvera compression for page-level encoding, and the organization of data into three distinct collections (Commercial Agreements, Corporate & IP Agreements, Operational Agreements) to facilitate agentic routing. The Query Agent's 'Search Mode' and 'Ask Mode' are showcased for discovery and answer synthesis, respectively, with results streamed back with cited source passages. The post also provides a detailed prompt for building a similar application using the CUAD dataset and outlines critical implementation notes for the Weaviate client, dependency injection, and the sources endpoint.
  10. Introducing Weaviate Agent Skills | WeaviateIntroduces Weaviate Agent Skills, a repository of Weaviate-specific scripts and project cookbooks to enable coding agents to interact with Weaviate infrastructure. The repository is structured into Weaviate Skills (for cluster management, data lifecycle, and agentic search) and Cookbooks (for end-to-end project blueprints like chatbots and RAG systems). It details commands for schema inspection, data import, and various search modes, and provides installation instructions for common coding agent tools.