Developer builds Sirro to test whether code memory survives switching AI agents

·1 views

A developer built a tool called Sirro to address a recurring problem: code saved in one AI coding agent cannot be reliably reused when switching to a different tool or project. The core challenge is distinguishing true retrieval of saved code from an agent simply regenerating something that looks similar. To test this, the developer proposes saving a working code snippet via one agent, then opening a separate tool in a fresh project and requesting the same asset by name. Sirro is currently in beta and has faced authentication issues, which the team is fixing before inviting broader user testing. The developer argues that a genuine agent-memory system must preserve specific implementation details across tool switches, not just make assets nominally available.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How One Developer Uses Traefik and Docker to Automate Reverse Proxy in a Homelab

A developer shared their homelab setup using Traefik as a reverse proxy to manage multiple Docker-based services such as Pi-hole, Vaultwarden, and custom applications. Rather than manually editing Nginx configs and juggling port numbers, Traefik automatically discovers Docker containers and routes traffic using container labels. The setup handles HTTPS certificates automatically via Let's Encrypt, eliminating the need for manual SSL configuration. Traefik is configured as the sole internet-facing service, while all other containers communicate internally through a dedicated Docker network. A key design choice is disabling automatic container exposure by default, ensuring only explicitly labeled services are made publicly accessible.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

OGAD App Lets Users Edit and Review Documents Offline Using Local AI Models

OGAD (Off Grid AI Desktop) is a desktop application for Mac, Windows, and Linux that enables users to review, outline, and draft documents using a locally run AI model without requiring an internet connection. The tool is designed for situations where connectivity is unavailable, such as during flights, allowing users to work through saved files in TXT, Markdown, DOCX, or PDF formats. Users must prepare the app, download a local language model, and save all necessary documents before going offline, as the AI cannot access current or external information. The application works by extracting text from attached documents and responding to user queries grounded in those sources, though users are advised to verify outputs against originals. The developers note that offline capability does not make the model's general knowledge current, and any facts requiring live data should be flagged for post-flight verification.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Dev Team Builds AI Support Agent With Persistent Memory to Eliminate Repeat Queries

A development team has built SupportMemory, an AI customer-support agent designed to remember details from past customer interactions across separate conversations. The agent uses a tool called Hindsight as its persistent memory layer, enabling it to both recall relevant information from previous sessions and retain new details for future use. Most existing AI support systems can only access context within a single conversation, forcing customers to repeat information each time they return. SupportMemory addresses this by retrieving only relevant stored memory rather than loading an entire conversation history, making responses faster and more personalised. The team demonstrated the difference by comparing a standard stateless agent against their memory-enabled agent using example customer interactions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

JuiceFS Enterprise 5.4 scales to 1 million clients on 100-billion-file systems

JuiceFS has released Enterprise Edition 5.4, building on version 5.3's support for 100-billion-file environments to now handle up to one million simultaneously connected clients. To achieve this, the team co-located metadata nodes with network proxies and introduced connection reuse, so that one million client connections were managed through just 10 proxy nodes with roughly 400 physical backend connections to the metadata leader. Session management was also overhauled using indexed lookups, paginated queries, and batch cleanup of expired sessions to reduce overhead at scale. The release additionally brings cross-zone directory cloning, over 90% RDMA bandwidth utilization, up to 100% improvement in high-concurrency random-read performance, and more flexible caching strategies. These enhancements are aimed at large-scale AI and enterprise workloads that demand both massive metadata capacity and high concurrent access.

0 comments Read more at DEV Community