An agent that remembers everything has bad memory — how to maintain useful memory
victorialslocum · x · 2026-08-26
This thread discusses memory design for AI agents. Core claim: an agent that remembers everything has bad memory — stale preferences, irrelevant tool outputs, and too many tasks competing for the model's attention. The author splits memory into two layers:
- Short-term memory: lives in the context window — active conversation, recent tool outputs, immediate reasoning state. Finite, and should stay lean.
- Long-term memory: stored externally (typically a vector database), retrieved when relevant; persists user preferences, past interactions, and domain knowledge across sessions.
Memory isn't just storage — it's maintenance. A write-only memory store becomes stale, contradictory, and expensive to search. Four operations keep it useful: pruning (removing outdated/irrelevant info), deduplication (merging duplicate entries), and more (thread truncated).
More from coding & agent
- Self-host a ChatGPT-style QA for team docs with one Docker command via kotaemon — KhuyenTran16 · 2026-08-26
- AI Agent Harness Hackathon underway: 5,000+ builders, DGX Spark prizes, 5 days left — KlausCodes · 2026-08-26
- Unify CTO breaks down how the team cut AI agent costs 95% in two weeks — LangChain · 2026-08-26
- Recursive Language Models & Prime Agent: A New Paradigm for Building Agents — CShorten30 · 2026-08-26
- Grok bot设想：通过X Money委托支付调用API链 — Daniel_Farinax · 2026-08-26
- DAIR AI Launches Interactive Tutorial for Building Agent Skills — dair_ai · 2026-08-26