Skip to main content

The Amnesiac Agent

Failing to manage what the agent remembers across turns and sessions. This manifests in two opposite ways: the agent accumulates unbounded context until quality degrades silently, or the agent starts each session with no memory of prior work, repeating effort and losing continuity.

Why It Happens​

  • Context management is unglamorous plumbing
  • Appending every message to the history is the simplest implementation
  • Summarization feels lossy — teams worry about dropping important details
  • The failure is gradual, not sudden, so it passes spot-checking
  • "Just use a bigger context window" feels like a solution

What Goes Wrong​

Unbounded accumulation:

No memory across sessions:

  • Context rot — LLM performance degrades with longer inputs even within the stated context window
  • Silent information loss — when context overflows, summarization drops details you didn't know mattered
  • Premature victory — agent sees partial progress, assumes work is done, and declares success
  • Escalating cost — unbounded context means every turn gets more expensive
  • Shift-handoff problem — cross-session agents are like engineers arriving with no memory of previous work

What to Do Instead​

  • Structured state — maintain progress in a structured format (checklist, state file) outside the context window
  • Strategic summarization — compress at natural breakpoints, not when the window overflows
  • Session initialization — start each session by reading progress logs, not by re-deriving state
  • Memory patterns — use appropriate memory patterns for the task duration
  • Sanity checks — verify state before acting (read the progress file, check what's already done)

Signs You Have This​

  • The agent occasionally "forgets" instructions from earlier in the conversation
  • Cross-session tasks start from scratch every time
  • You've never measured how context length affects output quality
  • There's no explicit state management — the conversation history is the state
  • Long conversations degrade in quality and you're not sure why