SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Claude Code memory: which files actually load at session start

0
·1 views

A test of ten memory setups on Claude Code version 2.1.300 measured how many tokens each configuration loads. Existing imported files and rules files without a path filter were found to load in full. An import pointing to a missing file loads nothing and produces no error, still ending with exit code 0. The tests also found that AGENTS.md beside CLAUDE.md is never read, and that imports inside path-filtered rules files load at session start regardless. Because every loaded character is charged on each request, these behaviors affect cost.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Developer builds offline Python tool using tiny local model to explain Swiss official letters

A developer grew weary of constant AI news and instead built a small private tool to explain Swiss administrative letters, such as payment orders and tax decisions, in plain Italian. Because these documents contain sensitive personal and financial details, the tool runs entirely on the user's machine using gemma3:1b through Ollama, with no API keys, accounts, or uploads. Plain Python code extracts exact figures like amounts and deadlines, while the model only paraphrases the explanation, so it cannot alter those values. The author reports that one run took about 40 seconds on CPU, a speed they consider adequate for this kind of task, though they note results will vary by hardware. They argue that matching a small model to a bounded job, rather than relying on large frontier systems, offers a more practical and less exhausting approach to AI.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Self-Hosting the prompts.chat Prompt Library Behind Traefik and Tailscale

A developer describes deploying f/prompts.chat, the open-source prompt library formerly known as Awesome ChatGPT Prompts, on a private homelab to avoid leaking internal prompts to third-party services. The setup uses Docker Compose with Traefik as a reverse proxy for automatic HTTPS, and restricts access through a Tailscale network with ACL rules and firewall changes that block public exposure. The author reports that concurrent edits caused SQLite lock errors, which were resolved by enabling Write-Ahead Logging, and that an early misconfiguration exposed the Traefik dashboard, which was indexed by Shodan within hours. The post lists performance figures such as sub-80ms warm responses and roughly 235MB combined memory use. It notes the approach is unsuitable for distributed teams without a VPN, multi-region replication, or strict compliance audit needs.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Free Browser-Based DodoDeck Offers PDF Editing, Text-to-Speech and File Sharing

A developer has launched DodoDeck, a website offering a collection of free tools that run directly in the browser. The suite includes a PDF text editor, AI-powered text-to-speech, a tool for resizing images to exact kilobyte sizes, and peer-to-peer file sharing of up to 1GB. According to the post, the tools require no sign-up and add no watermarks to output. The announcement was shared on DEV Community as a response to paywalls on simple web tasks. The platform's marketing describes its tools as private and lightweight, though the source provides no independent verification of these claims.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Debugging a Failed Webhook Locally by Capturing, Fixing, and Replaying Requests

A developer explains how to debug a failing webhook without triggering new provider events each time. The walkthrough uses the Hooklistener CLI to tunnel public HTTPS requests to a localhost app on port 3000, or to forward them from an existing Hooklistener endpoint. Developers capture a test delivery, inspect its method, path, headers, and body, then replay it against the fixed handler to confirm the response and side effects. The guide warns that replays keep the original event ID and that signature checks and timestamps can cause failures, so duplicate handling should be tested deliberately. Saving the captured request as a replay case lets developers rerun it after future changes.

0 comments Read more at DEV Community