Architecture
APR 2026
Composite scoring: fixing stale agent recall
Pure cosine similarity has no concept of time, so a four-month-old memory can overwrite the right context. The composite-scoring formula, anti-stale penalty, and 10-step retrieval pipeline that fixed it, over a 230K-vector store.
Architecture
APR 2026
Context engineering in practice: a 3-tier memory system, 230K vectors
The architecture, code patterns, and production numbers behind a 3-tier memory system for one AI agent: a 75-line core file, a 230K-vector LanceDB archival store, and self-teaching procedural memory. 6,370 jobs, $0.003/month embedding cost.
Architecture
MAR 2026
The memory system and composite scoring
How one production AI agent remembers across 5,252 jobs: a three-layer memory store, an LLM judge that gates writes, and a 197KB identity layer assembled at boot. Part two of two.