Posts
All the articles I've posted.
The AI Memory Funnel: Triaging and Crystallizing Agent Context
Updated: at 06:00 AMAI memory isn't just dumping logs into a vector DB. It's a lifecycle. We built a three-tier funnel in membox—Trace to Unit to Crystal—that automatically decays noise, consolidates facts, and crystallizes high-frequency context without ever explicitly deleting anything.
Decoupling LLM Extraction: Building an Async Ingestion Queue in SQLite
Updated: at 07:00 AMLLM extraction is brutally slow. We solved it without a message broker: a single SQLite table acts as a durable async queue, a transient worker process drains it, and a lease in the meta table guarantees only one worker runs at a time, with built-in crash recovery.
Measuring AI Memory: Fusing BM25 with Graph Traversal
Updated: at 06:00 AMWhy pure graph traversal fails on long passages and pure vector search loses logical links, and how fusing SQLite FTS5 BM25 with graph BFS plus a dynamic token budget knapsack cutoff hit 92.3% recall in our offline setting.
Concurrency in Local-First AI: Making SQLite Multi-Agent Safe
Updated: at 08:00 AMHow to fix SQLite database locks when multiple AI agents write to memory simultaneously, using WAL mode, threading.local, and application-layer RLocks.
Graph Traversal Without a Graph Database: BFS in Pure Python
Updated: at 07:00 AMWhy Neo4j is overkill for a personal AI agent's memory, and how to implement fast, multi-hop context retrieval using BFS in pure Python over SQLite.
Building a Deduplication Pipeline for Local Knowledge Graphs
Updated: at 06:00 AMWhy LLM extraction fails at knowledge graphs, and how a cascade of exact aliases, embeddings, and string fallback keeps local SQLite graphs clean.