Why AI Memory Systems Are Mostly Garbage

The Case Against Automated Memory

Modern coding agents are increasingly adopting 'memory' features that automatically save state to hidden files across sessions. The core argument against these systems is that they prioritize persistence over accuracy. In software development, the codebase itself is the ground truth. When agents attempt to maintain external memory files, they create a 'split-brain' problem where the agent relies on stale, context-heavy data that may have been true weeks ago but is actively harmful today. This is analogous to out-of-date code comments, which are worse than no comments at all because they mislead the developer and the model.

The 'Slop' Problem

An audit of Claude Code's memory files on a primary development machine revealed that out of 45 stored memories, 26 had never been read once. The stored data included transient information like specific PR numbers, temporary environment configurations, and point-in-time debugging logs that were no longer relevant. This 'slop' accumulates as dead weight, forcing the model to parse irrelevant context that can bias its reasoning. The data shows a 3-to-1 write-to-read ratio, proving that these systems are far more aggressive at polluting the environment than they are at providing utility.

The Superiority of Bash and Tooling

Instead of relying on opaque memory systems, the industry is shifting toward 'bash-first' agentic workflows. Modern models are sufficiently capable of using standard shell tools to explore a codebase, read files, and verify state on demand. By treating the file system as the primary source of truth, agents avoid the pitfalls of hallucinated or cached memory. If an agent needs persistent context, it should be explicitly managed in a single, human-readable file (like an agents.md file) that the developer can maintain, rather than allowing the agent to silently write to hidden directories.

When Memory Actually Works

Memory is only useful in non-code contexts where there is no direct, programmatic path to the truth. In general-purpose chat (like ChatGPT), memory can help bridge the gap between disparate user topics (e.g., connecting a question about posture to a previous purchase of a desk). However, in coding, the relationships between files and functions are deterministic. If an agent cannot find what it needs, it is a failure of the agent's tools, not a lack of memory.