Claude Code's auto memory feature found mostly empty across 21 real-world projects

·1 views

An audit of Claude Code's auto memory feature across 21 project directories on a single Mac revealed that 17 were completely empty, despite the feature being enabled by default. The remaining four directories contained two to four notes each, and three of those independently recorded the same correction — highlighting that memory does not sync across repositories. Auto memory, distinct from the user-written CLAUDE.md file, allows Claude to save session notes such as user preferences, feedback, project decisions, and references into per-repository directories under ~/.claude/projects/. The feature has a load limit of 200 lines or 25KB at conversation start, is machine-local, and is not shared with subagents. The author chose to disable auto memory in one high-stakes repository running an autonomous agent, citing concerns beyond context size.

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)

Self-Hosted AI-Git-Bot Offers Code Reviews Without Per-Seat SaaS Fees

AI-Git-Bot is an open-source, self-hosted bot that automates code reviews, unit test generation, issue triage, and documentation updates directly within Git platforms like Gitea. The tool addresses two common pain points with commercial AI code review services: per-developer pricing that scales with headcount rather than value, and the requirement to send code to third-party servers. The bot runs as two Docker containers and can operate entirely on local open-weight models via Ollama, eliminating API keys and per-token billing. A 7B-class model such as qwen2.5-coder:7b runs on 8 GB of RAM without a GPU, making it viable on a low-cost VPS. The Docker image has already been pulled over 16,000 times and supports both x86 and ARM64 architectures.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Three critical settings that prevent data loss when self-hosting AnythingLLM

Self-hosting AnythingLLM via Docker is straightforward, but redeployments silently wipe all workspaces, embeddings, and user data if three key settings are misconfigured. The first involves correctly setting the STORAGE_DIR environment variable and ensuring a persistent volume is actually mounted at that path, since an unmounted path causes irreversible data loss without any warning. The second is a file permissions issue, where a fresh volume mounted to a non-root container user will block writes and trigger misleading SQLite errors, requiring either root access or a pre-boot ownership change. Third, platforms like Railway and Heroku dynamically assign ports via a $PORT variable, but AnythingLLM defaults to port 3001, causing 502 errors unless PORT is explicitly pinned to 3001. The author, who maintains a Railway deployment template for AnythingLLM, notes these fixes apply regardless of whether the app is run on Docker, Fly, Render, Coolify, or similar platforms.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Live Data Shows 77% of Austrian Truck Parking Fills Up Overnight, Stranding Drivers

EU regulation 561/2006 mandates strict driving-hour limits for truck drivers but says nothing about whether parking is actually available when rest breaks are required. Analysis of real-time occupancy data from 85 Austrian motorway truck-parking sites — polled every 60 seconds since June 2026 — shows network-wide occupancy nearly doubles from around 38% at 11:00 to over 76% by 21:00, remaining at that plateau until roughly 4:00 a.m. Critically, the load is unevenly distributed: 40 of the 86 monitored sites recorded zero free bays at least once in the past 30 days, and 33 sites were completely full at some point during the 21:00 hour alone. A typical site like Wiener Neustadt, with 40 bays, averages fewer than 6 free spaces by 20:00 and has hit zero in every evening hour between 18:00 and 23:00. Static mapping tools continue to display full site capacity regardless of actual availability, leaving dispatchers and drivers without the real-time insight needed to plan legally compliant rest stops.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why One Developer Writes Two Separate Release Notes for Every Shipped Update

A developer at Burning Tribe has adopted a practice of writing two distinct release notes for every software update instead of a single shared changelog. The first note, aimed at users, is limited to three lines and describes only observable changes in plain language, omitting internal details like library versions or refactors. The second is a private note for the developer's future self, recording what problem was being solved, what approaches were tried and discarded, and what conditions might prompt revisiting the decision. Keeping the two notes separate prevents the common pitfall of writing for mixed audiences, which the developer says caused them to quietly abandon a single changelog over time. The author argues that drafting the user-facing note first also helps catch half-finished or bundled changes before deployment, making it a lightweight quality check.

0 comments Read more at DEV Community