Evolution of AI Agents: Memory, Collaboration, and Trust Challenges
에이전트 | Tue Jun 30 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 3 sources
Microsoft unveiled the Memora long-term memory system while debates intensified over deploying AI agents in the workplace and enterprise trust levels.
Analysis
[Microsoft Research] unveiled Memora long-term memory system [1]
- Achieved SOTA on LoCoMo and LongMemEval benchmarks
- Uses up to 98% fewer context tokens compared to Mem0
- RAG
- and full-context inference
- Harmonic memory architecture separating stored content from retrieval methods
- To be presented at ICML 2026 and open-sourced on GitHub
[MIT Technology Review] criticized the practice of referring to AI agents as colleagues [2]
- Cited research by Emma Wiles at Boston University
- Error detection rates dropped 18% when AI was framed as an employee
- About one-third of surveyed managers' companies frame AI agents as employees
- 23% list AI agents on their organizational charts
[Daron Acemoglu (MIT)] argued for agent design centered on complementing human capabilities [2]
- 2024 Nobel laureate in Economics
- Current marketing direction leans toward 'human replacement'
- Cited Stanford research on 1
- 500 workers
- Workers prefer automation only in specific areas such as legal assistants
[Gartner & McKinsey] projected 2026 as the 'inflection year' for enterprise AI [3]
- IT infrastructure costs expected to grow 2-3x by 2030
- Budgets remain frozen
- Technical teams are the key drivers of agent adoption
- Increasing pressure to prove ROI
[MIT Technology Review Insights] released findings from an agent trust survey of 300 technology professionals [3]
- Evaluated 101 tasks across AI
- data
- and cloud workflows
- Trust surges for measurable tasks
- Lack of business context is the weakness for more complex tasks
- Human oversight is the key factor for success