Field Notes: AI — May 2026
The month in one paragraph
May 2026 was the memory-hygiene month. Agent memory stopped being a sentimental feature and became an operational one. The useful question was not "does the agent remember?" It was "can the agent consolidate run history into better future execution without carrying stale, duplicated, misleading, or overbroad context?" Instruction files, run histories, skills, repo manifests, and persistent memory all pointed to the same design pressure: context has to be managed like infrastructure.
Actual field update
- Memory became operational: run history, project conventions, tool results, and user preferences became inputs into future execution.
- Stale context became a failure mode: more memory can degrade agents if authority, freshness, and scope are not tracked.
- Manifests became project interfaces: CLAUDE.md / AGENTS.md-style files encoded commands, architecture, and operational rules for agents.
- Skills/playbooks gained importance: reusable procedural knowledge started to look cleaner than spawning many specialized agents.
Robustness check
Strong claim: context and memory are now core agent infrastructure.
Moderate claim: manifests improve agent behavior when structured and current.
Weak claim to avoid: "more memory equals better agents." Poor memory increases false confidence and stale decisions.
Agentic design pattern change
The pattern became:
run trace
→ extract durable facts
→ classify freshness / authority / scope
→ consolidate memory
→ remove stale or duplicate entries
→ feed next run selectively
Memory is not storage. Memory is selective reuse.
Fallout
- Context freshness needed explicit metadata.
- Memory needed deletion and contradiction handling.
- Skills/playbooks became preferable to ad hoc prompt lore.
- Long-term memory created new privacy and governance concerns.
What builders should copy
- Store memory entries with source, timestamp, authority, scope, and last-use outcome.
- Separate project rules from user preferences and temporary task facts.
- Add stale-context detection.
- Track whether retrieved context helped or hurt the run.
- Prefer small procedural skills over giant global instructions.
Resource sources
- Agentic Coding Manifests / Claude.md study: https://arxiv.org/abs/2509.14744
- Claude Code design-space analysis: https://arxiv.org/abs/2604.14228
- Anthropic — Building Effective Agents: https://www.anthropic.com/engineering/building-effective-agents
- OpenAI Agents SDK — state, tools, orchestration: https://developers.openai.com/api/docs/guides/agents