AI 에이전트용 메모리를 구축하다가 이상한 점을 발견했습니다
We discovered something strange while building memory for AI agents
핵심 요약
AI 에이전트 메모리 구축 시 핵심은 저장보다 무엇을 기억하지 않을지 결정하는 '메모리 거버넌스'임을 발견했습니다.
- 메모리 거버넌스 — 정보 저장보다 불필요한 기억을 관리하는 것이 성능에 더 중요함
- 성능 저하 원인 — 에이전트 경험이 쌓일수록 검색 품질이 떨어지고 노이즈가 발생함
- 연구 과제 — 메모리 감쇠, 강화, 반추 품질 및 재실행 시스템 등을 탐구 중
- 오픈소스 프로젝트 — CogniCore를 통해 에이전트 메모리 인프라를 구축하고 기여자를 모집함
Over the last few months I've been building CogniCore, an open-source memory infrastructure layer for agents.
Initially, I assumed the hard problem was memory.
I was wrong.
The hard problem is deciding what NOT to remember.
We recently built:
- MCP server integration
- LangChain integration
- CrewAI integration
- OpenAI Agents SDK integration
- Episodic + semantic memory
- Reflection engine
- Replay and branching system
Then we started benchmarking.
The surprising result:
A simple memory system often performs almost as well as a much more sophisticated memory architecture.
The problem isn't storing information.
The problem is memory governance.
As memory grows:
- old strategies become stale
- failures get duplicated
- retrieval quality degrades
- reflections start conflicting
- agents become distracted by irrelevant experiences
At 10 episodes everything looks great.
At 500 episodes most memory systems become noisy retrieval systems.
This is the problem we're now trying to solve.
Current areas we're exploring:
- Memory decay
- Should memories expire?
- When?
- Memory reinforcement
- Should successful memories become stronger over time?
- Reflection quality
- Why do some reflection systems improve performance while others make agents worse?
- Replay systems
- Can agents learn from previous trajectories without retraining?
- MCP-native memory
- What does persistent memory look like when every tool can access it?
We're looking for contributors interested in:
- Agent architectures
