MongoDB ships VFS backend for LangChain Deep Agents: S3 storage, Atlas search
A new `langchain-mongodb-deepagents-vfs` backend implements the `BackendProtocol` interface in LangChain Deep Agents, routing file operations across two stores. Reads, writes, and edits go to S3 as the source of truth, while listing, globbing, and grepping run against MongoDB Atlas metadata, chunks, and embeddings. Practitioners can change storage and search backends without touching agent code, prompts, or subagent wiring.
- langchain-mongodb-deepagents-vfs routes read/write/edit to S3 (source of truth) and ls/glob/grep to MongoDB Atlas, storing path metadata, chunks, and embeddings
- grep fuses full-text and vector search via $rankFusion, returning line-oriented results for literal and semantic queries
- Backend swaps via the BackendProtocol interface without changing agent code, prompts, or subagent wiring; code is on the langchain-ai GitHub repo
Related
Developer argues file systems are the most intuitive data interface for AI agents
A developer explores whether file systems should serve as the primary data interface for AI agents needing persistent, retrievable, modifiable state across sessions. The core argument: LLMs already handle Unix file operations fluently from training data, so exposing data as files avoids building custom interfaces per system. The idea is framed as an open exploration, not a shipped tool or benchmark.
A developer explores whether file systems should serve as the primary data interface for AI agents needing persistent, retrievable, modifiable state across sessions. The core argument: LLMs already handle Unix file operations fluently from training data, so exposing data as files avoids building custom interfaces per system. The idea is framed as an open exploration, not a shipped tool or benchmark.
Developer argues file systems are the natural data interface for AI agents
A developer explores whether file systems should serve as the primary persistent data interface for AI agents, arguing files beat databases, APIs, and object storage because LLMs already deeply understand Unix file operations. The idea matters to agent builders choosing state-management primitives for multi-session agents.
A developer explores whether file systems should serve as the primary persistent data interface for AI agents, arguing files beat databases, APIs, and object storage because LLMs already deeply understand Unix file operations. The idea matters to agent builders choosing state-management primitives for multi-session agents.
Hugging Face engineer details Hub search scaling for 3M models
In a YouTube talk, Hugging Face engineer Arek Borucki explains how the Hub keeps search instant across 3 million public models, 14 million users, and a million datasets. The system uses Apache Lucene full-text search via MongoDB Atlas, with model artifacts stored in S3 for independent compute scaling.
In a YouTube talk, Hugging Face engineer Arek Borucki explains how the Hub keeps search instant across 3 million public models, 14 million users, and a million datasets. The system uses Apache Lucene full-text search via MongoDB Atlas, with model artifacts stored in S3 for independent compute scaling.
Developer shares agent-as-memory-controller architecture using Neon Postgres
A developer describes their long-term memory architecture for AI agents, using an agent-as-memory-controller pattern with hierarchical directory memories stored in Neon Postgres. The approach uses tools for CRUD operations and leverages Neon's scale-to-zero and instant branching for debugging.
A developer describes their long-term memory architecture for AI agents, using an agent-as-memory-controller pattern with hierarchical directory memories stored in Neon Postgres. The approach uses tools for CRUD operations and leverages Neon's scale-to-zero and instant branching for debugging.
First systematic study of filesystem-based memory for LLM agents
A new paper presents the first systematic exploration of filesystem-based memory for LLM agents, where agents store long-term memory as a directory tree of markdown files they read, write, and reorganize via generic file tools. The study tests two untested assumptions: that agents can keep a growing store organized as memories accumulate, conflict, and go stale, and that this organization pays off.
A new paper presents the first systematic exploration of filesystem-based memory for LLM agents, where agents store long-term memory as a directory tree of markdown files they read, write, and reorganize via generic file tools. The study tests two untested assumptions: that agents can keep a growing store organized as memories accumulate, conflict, and go stale, and that this organization pays off.