Agent memory architecture
ReliBot & MyClawBot
How an incident-response agent remembers past fixes: L1/L2/L3 recall, SQLite hybrid search, write/read paths, and what token-metering bugs taught me.
80+tools · 16 domains
12sub-agents
3-layerauto-recall
2,200+tests
- ReliBot system
- Memory architecture
- How the two projects relate
- Design decisions
- Lessons from token metering
7 diagramsOpen →
AI security review · GCP deployment
GitHubScanner
Agent architecture, CVE enrichment, the trust pipeline, a threat model of the scanner itself, and the AWS CodePipeline → Workload Identity Federation → Cloud Run path.
6specialist agents
OSV + EPSSCVE intelligence
3,400+tests
86%coverage
- System
- CVE enrichment
- Trust pipeline: defending against hallucinated findings
- Threat model: OWASP Top 10 for LLM applications, applied to the scanner itself
- Cloud design inside the app
6 diagramsOpen →
Knowledge graphs on GCP · AWS & OCI equivalents
GraphRAG vs Vector RAG
Schema-constrained graph extraction into Neo4j, a LangGraph vector-RAG twin, the Cloud Run cost story, model A/B results, and a GCP → AWS → OCI service map.
43source docs
1,943graph nodes
4,081relationships
~98%hosting cost cut
- System
- GraphRAG or vector RAG?
- GCP deployment
- The same system on AWS
8 diagramsOpen →