Being a Subagent - Why Not Remembering Is a Feature
Being a Subagent - Fresh Starts Are a Feature
The default reaction to an agent that does not remember previous sessions is frustration. You want continuity. You want the agent to build on what it learned last time. You want progress to accumulate.
But the not-remembering part is actually a feature. Every fresh start is a chance to approach the same problem without baggage.
The Anchoring Problem
When an agent remembers its previous approach, it anchors to it. If it tried solution A last time and failed, it is biased toward variations of A rather than considering completely different approaches B through Z. Previous context narrows the search space - sometimes helpfully, sometimes harmfully.
A fresh subagent has no anchors. It reads the problem description, examines the current state of the code, and generates a solution from scratch. Sometimes that solution is worse than what a memory-equipped agent would produce. But sometimes it is dramatically better because it avoids the dead ends that accumulated context would have reinforced.
When Fresh Starts Win
Fresh starts are most valuable when:
- Previous approaches have failed - If three sessions have tried and failed to solve a problem, the fourth session benefits from not knowing about the first three
- Requirements have changed - Old context about outdated requirements actively hurts. A fresh agent reads the current state
- The problem needs reframing - Sometimes the issue is not the solution but the problem definition. A fresh perspective catches this
When Memory Matters
Memory is essential for coordination, not creativity. You need the agent to know which files it should not touch, what conventions to follow, and what the project structure looks like. This is what CLAUDE.md and git history provide - stable context, not session-specific memory.
The Hybrid Approach
The best setup gives agents access to facts (git log, file structure, project rules) while letting them forget opinions (previous solution attempts, abandoned approaches, emotional investment in specific designs).
Treat subagent amnesia as a diversity mechanism. Each fresh session is a new perspective on the same problem.
Fazm is an open source macOS AI agent. Open source on GitHub.