Developers Build Proxy Tool to Give CLI AI Agents Persistent Project Memory

·1 views

A team called AI Alchemists built cliMEM, a lightweight proxy tool, during the WeMakeDevs hangover hackathon to solve a recurring frustration: CLI coding agents losing all project context when a terminal session ends. The tool sits between the user's CLI agent and their AI provider, logging conversations and extracting key facts — such as decisions, conventions, and unresolved threads — at session end using rule-based heuristics. These facts are stored in Cognee, a graph and vector memory engine scoped per project directory, and automatically injected into the system prompt at the start of each new session. Setup requires just two commands, and the tool supports agents like Claude, OpenCode, and Codex without altering their original configurations. During development, the team also debugged a tricky shutdown crash caused by Cognee's database schema not being initialized when a session produced no memorable facts.

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)

PostgreSQL Replans Prepared Statements When Planner Settings Change at EXECUTE

PostgreSQL offers planner configuration parameters like enable_seqscan and enable_indexscan that shape how query execution plans are generated, not how already-planned queries run. A key question arises with prepared statements: do these settings take effect at PREPARE time, at EXECUTE time, or both? Testing with a simple tasks table and a partial index shows that when enable_indexscan is turned off before EXECUTE, PostgreSQL generates a new plan using a sequential scan instead of the previously chosen index scan. This means planner settings applied before EXECUTE can trigger replanning, overriding the plan cached at PREPARE time. Developers should be aware that changing planner parameters between prepare and execute can significantly alter query behavior and performance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Dev builds internal-linking system to help new tool pages rank on Google

A solo developer building an image-processing tool in public documented week 17 of their project, focusing not on new features but on SEO infrastructure. After shipping an AI background remover the previous week, they discovered the new tool's pages were orphaned — reachable via sitemap but not linked from anywhere on the site. To fix this, they replaced scattered hardcoded link lists with a single source-of-truth constant that automatically cross-links a hub page and its use-case spoke pages. The same constant also drives layout logic and populates a footer column, meaning any future spoke page requires just one line of code to be fully integrated. A React hydration error caused by invalid HTML nesting was also caught and resolved during the process.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Omnikon Org's AI-Powered Developer Browser CNTRL Accepted Into ECSoC 2026

Omnikon Org's open-source project CNTRL has been selected for the Elite Coders Summer of Code (ECSoC) 2026 program. CNTRL is an AI-first browser designed specifically for developers, aiming to integrate AI assistance, context-aware documentation, and streamlined research workflows directly into the browsing experience. The selection provides the team an opportunity to collaborate with contributors worldwide and accelerate the project's development. Built using technologies including React, TypeScript, Tauri, and Rust, the project is open to contributions from developers of all skill levels. Omnikon Org, which focuses on creating modern developer tools, says ECSoC participation marks a significant milestone in realizing its vision of making open-source development more accessible and impactful.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Add Security Guardrails to Python LLM Apps Against Prompt Injection

A technical guide published on DEV Community outlines key security boundaries in large language model (LLM) applications built with Python. The article identifies three entry points for untrusted text: user input, retrieved content from RAG pipelines, and model output that feeds into further processing. It focuses on defending against prompt injection attacks, where malicious text attempts to override a model's instructions, by keeping user input in clearly delimited data channels separate from system instructions. The guide also covers input validation, output validation, and PII redaction as part of a broader 'security-forward' approach to LLM app development. Code examples using Anthropic's API illustrate the difference between unsafe string concatenation and safer, delimiter-based prompt construction.

0 comments Read more at DEV Community