Jean2 AI Coding Agent Builds Persistent Memory Across Sessions via Local Files

·1 views

A developer frustrated by AI coding assistants forgetting project context after each session built Jean2, an agent designed to retain knowledge across sessions without retraining or vector databases. Jean2 stores and reads project-specific information in plain markdown files on disk, allowing it to remember details like build tools, database choices, and folder restrictions. The system uses four opt-in knowledge layers — workspace memory, agent memory, workspace skills, and agent skills — each scoped either to a project or to the individual agent. Users must enable these layers manually in workspace settings, meaning the agent starts blank by default and is customized incrementally. The approach lets the agent behave more like a knowledgeable colleague, accumulating corrections and preferences over time rather than repeating the same onboarding every session.

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 Much Observability Do Microservices Actually Need to Be Worth the Cost

A technical discussion on DEV Community explores the real-world trade-offs of implementing observability in microservice architectures. Unlike basic monitoring, which signals that a problem exists, full observability combines metrics, logs, and distributed tracing to pinpoint exactly where and why failures occur across multiple services. The author illustrates that debugging a multi-service outage without tracing can take over an hour, while proper tracing can reduce that to under two minutes. However, the costs are significant: a 20-service system using 100% trace sampling can generate roughly 4 terabytes of data per day, with monthly storage bills ranging from $90 on AWS S3 to over $2,000 on Datadog. The piece advises teams to balance observability depth against budget and team capacity, recommending strategies like tail-based and error-priority sampling to manage costs without sacrificing critical visibility.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Spec-Driven Development gains traction as AI coding tools create documentation gaps

A growing movement in software engineering called Spec-Driven Development (SDD) is pushing developers to write detailed specifications before writing any code. The approach addresses a key problem emerging from widespread AI coding tools like GitHub Copilot and Cursor: projects increasingly lack clear documentation of why decisions were made or how components interact. Instead of feeding vague prompts to AI, developers write structured specs that serve as a single source of truth, from which AI agents can consistently generate architecture, tasks, and code. GitHub recently released an open-source project called Spec Kit to formalize this workflow. Proponents argue that while AI has made code generation trivially easy, clearly defining what needs to be built remains the hardest and most valuable part of software development.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Pick the Right Headless CMS: Key Factors Developers Should Consider

A headless CMS decouples content editing from content display, delivering content to any front end via API, but choosing the wrong one can frustrate editors or create fragile integrations. The main categories include API-first platforms like Sanity and Contentful, git-based systems suited to developer-managed content, and visual page-builder hybrids like Storyblok that appeal to non-technical teams. Critical evaluation factors include content modeling flexibility, editor experience, developer tooling, pricing structure, and how difficult it is to migrate away from the platform. Experts caution that a headless CMS is best suited for sites with active editorial teams, while structured application data is often better managed with a standard database and lightweight admin interface. Getting the decision wrong is costly to reverse once a content team has built workflows around a particular tool.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How JVM Memory Behavior Outside the Heap Can Crash Kubernetes Pods

An iPaaS platform running on Kubernetes experienced repeated pod restarts despite healthy heap usage and stable application performance. The root cause was total JVM process memory — including Metaspace, thread stacks, direct buffers, and GC metadata — exceeding the container's memory limit. Because Kubernetes enforces limits on total process memory rather than heap alone, the Linux kernel terminated the JVM process each time it breached the threshold. The incident underscores a common blind spot for developers who treat the JVM as a black box while focusing on integration logic and APIs. As containerized deployments impose tighter resource constraints, understanding JVM internals such as memory layout, classloading, and GC behavior has become critical for building reliable microservices.

0 comments Read more at DEV Community