Developer builds local MCP memory layer to cut AI context costs in coding sessions

·1 views

A developer has built an open-source tool called zerikai_memory to solve the recurring problem of re-establishing project context at the start of every AI coding session, which wastes tokens and increases API costs. The tool runs as a memory layer behind Claude Desktop via the Model Context Protocol (MCP), storing indexed project knowledge locally rather than relying on the AI's context window. It uses Tree-Sitter to parse source files into Concrete Syntax Trees without any LLM inference, supporting Python, JavaScript, TypeScript, HTML, and Markdown. A multi-stage retrieval pipeline — combining vector search, lexical re-ranking, and source citations — ensures only the most relevant context is surfaced to the IDE or chat session. The system supports fully local operation via Ollama, cloud mode via DeepSeek, or a hybrid setup, and is designed to keep API bills manageable while preserving session continuity.

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)

Why AI Agents Need Stricter Architecture Decision Records Than Humans Do

A senior software engineer and tech lead building toward fully autonomous 'Level 5' AI development has published a detailed framework for writing Architecture Decision Records (ADRs) specifically designed for AI agents. Unlike human-readable ADRs, agent-readable versions require four additional sections: an invariant statement, a list of 'dangerous improvements,' yes/no agent check questions, and a consequence table of observable violations. The need for this stricter format was demonstrated through a real codebase incident, where an AI agent optimizing an order creation endpoint made a locally reasonable change that broke a critical system invariant. The author credits Dan Shapiro's 'Five Levels' framework and AI strategist Nate B. Jones for the conceptual foundation behind the newsletter series. The core argument is that agents cannot infer intent or ask clarifying questions mid-implementation, making explicit, machine-legible constraints essential before violations are embedded in code.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Diagnose and Fix Next.js Performance Issues the Right Way

Developers often approach Next.js optimization too broadly, making multiple changes at once without knowing which one actually helped. A more effective method involves identifying a specific slow route and symptom before touching any code. Key diagnostic steps include recording baseline metrics such as TTFB, LCP, bundle size, and request count, then determining whether the bottleneck lies on the server or in the browser. Tools like bundle analyzers and simple timing helpers can pinpoint whether the issue stems from a slow CMS, oversized client components, or third-party scripts. The guide also recommends keeping components server-side by default and only using client components where browser interaction is genuinely required.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Refactor a Bloated Next.js Middleware File Using the NEMO Pattern

Next.js middleware files often start simple but grow unwieldy as redirects, locale handling, authentication, and security headers accumulate in a single file. The core problem is mixed responsibility, where unrelated logic coexists without a clear execution order, making changes risky and hard to test. A practical solution involves splitting middleware into discrete rules, each handling one concern, arranged in an explicit priority sequence. The NEMO library formalizes this composition pattern, treating middleware as an ordered set of independent proxy rules rather than a chain of nested conditionals. Keeping internal paths like static assets and health checks at the top of the rule list reduces unnecessary processing and prevents unintended side effects.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Six Private Dart/Flutter Package Registries Compared: Pricing, Features, and Trade-offs

A detailed comparison of six private package registry providers for Dart and Flutter — Cloudsmith, Inedo ProGet, JFrog Artifactory, OnePub, GLPub.dev, and Publy — has been published, with pricing verified against each provider's own pages as of August 2026. Options range from cloud-hosted general-purpose registries like Cloudsmith (free tier: 500MB storage) to self-hosted solutions like ProGet, which offers unlimited users on its free edition. Dart 3.9's new tag_pattern feature for git dependencies reduces the need for a registry in smaller projects, but limitations around mutable tags, URL-based package identity, and per-repo access management persist. For teams managing a handful of packages, git-based workflows with tag_pattern may suffice, but larger interdependent multi-package setups are likely to benefit from a dedicated registry. Billing models vary significantly, spanning per-seat, flat per-organisation, usage-based, and flat per-server licensing structures.

0 comments Read more at DEV Community