Auto Memory vs CLAUDE.md: Claude Code Reliability Tested Across Fresh Sessions

·1 views

A developer experiment tested whether a single fact stored in Claude Code's auto memory system could reliably guide behavior in new sessions, the same way a CLAUDE.md instruction does. Across four fresh sessions, the auto memory line succeeded on the first attempt in three cases and after one retry in the fourth, while the CLAUDE.md line succeeded first-attempt in all four sessions. Sessions with neither mechanism in place failed to recover in any of the four runs, tested on Claude Code version 2.1.278. The experiment was prompted by a reader's challenge to actually measure whether captured memories influence later sessions, rather than assuming they do. Both memory systems load text at session start as context rather than enforced configuration, meaning neither mechanism guarantees behavior — a distinction the researchers noted as key to interpreting the results.

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)

Developer builds bare-metal x86 OS using Claude AI, directing every line of code

A hobbyist developer has created BooleOS, a 32-bit x86 operating system written in C99, without personally writing any of the code — instead directing Anthropic's Claude Code AI agent and reviewing every change. The project targets real hardware and enforces strict constraints, such as fixed-width data types and a custom heap wrapper, to simulate the limitations of an actual kernel environment. The developer deliberately avoids standard tools like malloc and pthreads to prevent the AI from masking bugs that would only surface on the OS's eventual cooperative scheduling model. With around 20 tagged versions already released, the project is still pre-1.0 but represents a structured experiment in AI-assisted low-level systems development. The author highlights that reviewing AI-generated OS code demands a fundamentally different level of scrutiny compared to typical AI-assisted app development.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Custom Personal LifeOS to Log Tasks, Health, and Projects

A developer shared on DEV Community how they built a personal productivity system called LifeOS after finding existing tools like ChatGPT reminders and Obsidian templates too limited or cumbersome. The system is designed to handle quick daytime logging — such as meals, tasks, and project updates — while reserving heavier processing like daily reviews and project file updates for the evening. Data is split between a SQLite event log for real-time captures and a local Vault that holds the lasting state of projects, knowledge notes, and reports. The setup integrates with TickTick for daily task management and uses n8n for scheduled automation and webhook triggers. The developer's goal was a lean, adaptable system that keeps personal data on a home server while remaining accessible from both phone and work computer.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Free Tool to Help Students Meet Exam Form Photo and File Requirements

A developer has launched FormSaathi, a free browser-based platform offering nine tools designed to simplify the technical requirements of Indian competitive exam application forms. The tool addresses common pain points such as resizing photos to specific kilobyte ranges and pixel dimensions required by exams like SSC, UPSC, NEET, JEE, and others. All file processing happens locally in the browser, meaning no data is uploaded to external servers, which benefits students on low-end devices with limited internet access. The platform is also installable as a Progressive Web App and can function offline. FormSaathi is currently free with no sign-up or advertisements, and the developer is seeking user feedback to guide future improvements.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How SQL JOIN Fan-Outs Silently Inflate Your Totals and How to Fix Them

A common but hard-to-spot SQL mistake occurs when joining multiple tables causes row duplication, inflating SUM() and COUNT() results without triggering any error. This happens because a JOIN produces every matching row combination, so an order joined to four line items appears four times in the result, causing its amount to be counted four times. The phenomenon, known as a 'fan-out,' is especially deceptive because the query is syntactically valid and executes without warnings. Developers can detect the issue by running the JOIN without aggregates first and manually verifying row counts against known data. The recommended fix is to pre-aggregate the one-to-many table in a subquery before joining, ensuring each entity appears only once in the final result set.

0 comments Read more at DEV Community