What Does Remember Mean for an Agent? Store Everything, Prune 80%
What Does Remember Mean for an Agent? Store Everything, Prune 80%
What does "remember" mean for an agent? We took the literal approach first - store everything. Every interaction, every decision, every observation. After three weeks we had a massive memory store and an agent that was slower and less accurate than when it started.
The Accumulation Problem
More memory means more context. More context means more tokens. More tokens means higher latency, higher cost, and paradoxically, worse reasoning. The model spends attention on irrelevant memories instead of focusing on the current task.
An agent that remembers your coffee order, your file naming convention, your preferred git branch strategy, and also 47 debugging sessions from two weeks ago is not a better agent. It is a distracted one.
The 80% Prune
We pruned 80% of stored memories. The criteria were simple: when was this last relevant? Did it change the agent's behavior? Would the agent produce a different output without it?
Most memories failed all three tests. The debugging sessions were one-time events. The specific file paths changed. The meeting notes were never referenced again. What remained was durable knowledge - preferences, patterns, project structure, recurring workflows.
Responses Got Sharper
After pruning, response quality improved measurably. The agent was faster because context was shorter. It was more accurate because the remaining context was all signal, no noise. It felt more intelligent even though we removed information.
This is counterintuitive. We expect that more knowledge equals better performance. But for agents operating within context windows, the relationship is inverted past a threshold. Less relevant context beats more total context.
The Right Memory Strategy
Store everything initially - you do not know what matters yet. Review weekly. Prune aggressively. Keep only what changes behavior. Treat memory as a curated collection, not a database dump.
Fazm is an open source macOS AI agent. Open source on GitHub.