Contextual Relevance vs Over-Reliance: Managing 200 Lines of AI Memory
Contextual Relevance vs Over-Reliance: Managing 200 Lines of AI Memory
After maintaining a 200-line MEMORY.md file for months, I noticed something counterintuitive. More context does not mean better results. In fact, the relationship between memory size and agent performance follows an inverted curve - helpful up to a point, then actively harmful.
The Problem with Raw Context
Most developers start by dumping everything into their agent's memory file. Every API quirk, every workaround, every preference. The file grows to 500 lines and suddenly the agent starts applying advice from one project to a completely different one. It read that you prefer curl over fetch for one specific API, and now it uses curl everywhere.
Curated Pointers Beat Exhaustive Lists
The fix is treating MEMORY.md like an index, not a database. Each entry should be a pointer - just enough context to trigger the right behavior, not enough to override judgment. Instead of documenting every edge case of an API, write "Resend API - use curl via subprocess, not urllib (Cloudflare blocks Python UA)." That is the entire entry. The agent can figure out the rest.
Signs of Over-Reliance
Watch for these patterns:
- The agent quotes your memory file back to you instead of reasoning about the current task
- Old workarounds get applied to situations where the original bug was already fixed
- The agent follows memory instructions even when they contradict what you just asked for
The Retention Curve
I tracked which memory entries actually got used over 30 days. About 40% were never referenced. Another 30% were referenced but led to worse outcomes because the context had shifted. Only 30% consistently improved results. The lesson - prune aggressively and regenerate from recent data.
Practical Guidelines
Keep your memory file under 200 lines. Review it weekly. Delete anything older than 60 days unless it still applies. Group entries by project so the agent can skip irrelevant sections. And most importantly - write entries that guide judgment rather than dictate behavior.
The goal is an agent that uses memory as one input among many, not an agent that treats every memory entry as a commandment.
Fazm is an open source macOS AI agent. Open source on GitHub.