Adding Project Memory to Claude Code Cut Its Error Rate from 52.5% to Zero

·1 views

A developer tested Claude Code on 40 tasks drawn from real project history, comparing a standard session against one augmented with a curated memory tool called RE-call. Without memory, the AI made errors on 52.5% of runs; with memory, it recorded zero failures across all 40 paired trials. Answer correctness and factual accuracy both roughly doubled, and the memory-enhanced version never performed worse than the baseline in any single run. Each task contained a hidden pitfall — such as a command that appears to succeed but does nothing — that could only be avoided with knowledge of the project's past decisions. The experiment used deterministic checkers and repeated each task ten times per condition to account for the non-deterministic nature of AI agents, with p-values at or below 0.0003.

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)

Keybound tool audits prompt cache isolation in multi-tenant LLM relays

A developer named Pedro Sordo Martínez has released Keybound, an open-source auditing tool designed to verify prompt cache isolation in multi-tenant LLM relay systems. The tool tests whether a relay correctly prevents one tenant from reading cache data written by another tenant sharing the same upstream credential. Keybound implements the defense contract defined in the KeyPooling paper (arXiv:2608.17485), which requires that a namespace derived from authenticated tenant identity be applied to every cache lookup and write. The tool runs formal test fixtures against a synthetic mock gateway — returning a FAIL verdict when namespaces collapse and a PASS when per-tenant isolation is enforced. The project achieves 90% overall test coverage, with 97–100% coverage on the core audit logic, and is available under the AGPL-3.0-or-later license on GitHub.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Vagrant and Ansible Still Viable for Dev VMs, Though Containers Now Dominate

A 2014 blog post on setting up development virtual machines using Vagrant and Ansible has been reposted to DEV Community with minor edits. Vagrant was once the standard tool for spinning up local dev VMs, allowing multi-machine environments to be created in seconds. Unlike provisioning tools such as Puppet or Ansible, Vagrant itself only manages the VM lifecycle and delegates configuration to external tools. The author notes that while most teams now prefer containers, the Vagrant-plus-Ansible workflow remains relevant for use cases requiring full VMs, such as kernel-level testing or OS-specific work. An example project, vagrant-lamp-ansible, is still available on GitHub for developers who want to try the setup.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DEV Community Writer Explains Switching Profile Photo to VTuber Character Shigure Ui

A developer and writer on DEV Community recently changed their profile picture to an image of Shigure Ui, a character from the VTuber world, prompting them to write a detailed personal post explaining the decision. The author connects the change to themes explored in earlier writings about poverty, creative struggle, and emotional hardship. Growing up without internet or a computer in a small village, they describe having built their skills entirely from scratch while facing repeated professional setbacks. The profile picture change is framed not as trivial but as a meaningful reflection of where they currently find joy and emotional grounding. The post is intended as an honest account of how unexpected sources of comfort, including animated characters, can carry genuine personal significance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Custom New Relic Transaction Naming for eZ Publish via PHP Extension

A developer originally published in 2013 a solution for improving New Relic monitoring on eZ Publish PHP applications. By default, the New Relic PHP agent groups all eZ Publish transactions under '/index.php', making performance diagnosis difficult. The developer built an open-source extension called 'ezpublish-newrelic' that hooks into New Relic's PHP API to assign distinct, meaningful names to individual module and view transactions. The extension remains available on GitHub and can be installed via Composer from Packagist. Although New Relic's PHP agent has evolved since, the core approach of manually naming transactions instead of consolidating them into a single bucket remains a valid practice today.

0 comments Read more at DEV Community