Strategic Forgetting

Canonical version: Strategic Forgetting.

Strategic forgetting is choosing what an AI agent should forget, and when. It works at two time scales:

1. Inside a task: clear the context window on purpose, keep the artifacts, start fresh. This is the fix for Context Inertia
2. Across months: let long-term memory decay, merge or get rewritten, so that old memories don't pile up into contradictions

The first one is a known technique. The second one is, by most accounts, unsolved.

Forgetting inside a task

Research in 2026 keeps finding the same thing: agents that periodically lose their history do better than agents that keep it. On one optimization task (AHC014), a team of 4 agents that cleared its context while keeping its outputs scored 1,144,695, against 1,007,073 for 4 agents without forgetting (Shangyin Tan, UC Berkeley). Anthropic's Memento setup clears the context every ~90 minutes after the agent writes itself a note.

Practitioners do the same by hand: checkpoint the work into code, design docs and plans, then start a fresh agent from those. One HN commenter noted that design docs have become an output of coding as much as an input.

Forgetting across months

This is the hard part. Drew Breunig ran a Hermes Agent for six months and saw two things build up: self-written skills with contradictory fixes, and memories that conflict. His prediction: "overly precise memories become conflicted and messy", and memory decay is the next big problem. LangChain's Harrison Chase lists memory as an open problem for the same reason: storing memories is easy (a file, a string), but deciding what to remember is hard, and agents fail both ways, remembering nothing or every tiny detail.

Old facts rarely get deleted. They get contradicted by newer ones, and the model picks whichever version sits closer to the prompt. That's Context Clash built into the memory.

Approaches people are trying:

• Append-only plus gardening: never edit memories during work, then consolidate them periodically, like a Zettelkasten review
• Provenance on every fact: record where and when each memory came from, so a newer source can win explicitly (Context Provenance)
• Strike-through instead of delete: keep the old entry visible but marked outdated. That's what my own agents do in their MEMORY files
• Human review: periodically read the memory and prune it. Slow, but it works

Why I care

This is Knowledge Decay and Periodic reviews again, applied to machines. Personal knowledge management solved part of this decades ago: notes need maintenance, and a knowledge base that only grows gets worse. The same is true for an agent's memory. I expect the best agent memory systems to borrow more from PKM than from databases. Luhmann even saw forgetting as a feature of his Zettelkasten: forgetting leads to productive re-encounters.

References

• Drew Breunig, "The State of Context Engineering" (2026-10-24): https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html
• Harrison Chase, "New context engineering challenges in late 2026" (2026-11-02): https://www.langchain.com/blog/context-engineering-challenges-late-2026
• Hacker News discussion: https://news.ycombinator.com/item?id=50235780

Related

• AI Agent Memory
• Context Inertia
• Context Engineering
• Knowledge Decay
• Periodic reviews
• Context Provenance
• Context Clash
• AI Context Rot
• Hermes Agent
• Zettelkasten method
• The Zettelkasten is a tool to outsource memory, but also to enhance that fortuitous forgetfulness that leads to future productive re-encounters

About Sébastien

Ready to get to the next level?