Building Self-Correcting Systems: tools and research for AI agent memory reliability, authority auditing, freshness checks, and action-safe automation.
Current focus:
Relevance is not authority.
AI systems can retrieve context that is semantically relevant but not allowed to govern an action. My work explores how agent memory should track authority, freshness, scope, corrections, and verification requirements before tool use.
- 27 public research claims on AI memory reliability and authority boundaries.
- A live Memory Authority Auditor built with six specialized agent services.
- A public research harness with claim ledger, preregistrations, evaluators, ablations, and validity notes.
- Recent work on signed-and-fresh grants, paired authority/action logs, and scope- soundness boundaries.
A deployed multi-agent web app that audits AGENTS.md, CLAUDE.md, Cursor rules, SOPs,
and project memory files for stale instructions, authority conflicts, and verification
gates.
- Live app: https://memory-authority-auditor-web-992750435781.us-central1.run.app
- DEV submission: https://dev.to/zep1997/i-built-a-multi-agent-authority-auditor-for-ai-memory-files-1hb0
- Architecture: one Cloud Run web service plus six specialized agent services.
Agent roles:
- Memory Extractor
- Authority Classifier
- Conflict Detector
- Verification Gate Agent
- Authority Mapper
- Report Writer
A public research harness for testing whether memory retrieval systems select the memory that is authorized to govern an action, not only the memory that is most relevant.
- Repo: https://github.com/keniel13-ui/ai-memory-judgment-demo
- Includes lexical retrieval tests, embedding comparison, role-filter experiments, scope metadata tests, action-type arbitration, freshness gates, paired action logs, claim ledger, validity threats, and audit materials.
I write in public on DEV about AI memory, correction memory, uncertainty, authority arbitration, and production agent reliability.
- DEV profile: https://dev.to/zep1997
- Start here: https://dev.to/zep1997/start-here-my-ai-memory-research-so-far-4m4k
Production agent memory needs more than recall. Once agents can remember, use tools, and act across time, they need a trust layer around memory itself.
It needs:
- authority labels
- status and freshness
- scope boundaries
- source-of-truth pointers
- verification gates
- audit traces
- human review for sensitive actions
Memory is input.
Authority is the action boundary.