The Memory Problem — 20-post SilicaDB blog series (drafts)

[shiva]

Overview

Draft series: The Memory Problem — 20 posts building from first principles of AI agent memory to the introduction of SilicaDB.

Structure

• posts/ — draft markdown posts (Atlantic/New Yorker register, 800–5000 words, one key point each)
• research/ — per-post source notes with links for deeper exploration

Storytelling Framework

Each post uses one tactic from Storyteller Tactics (Pip Decks) — Dragon & The City, Man in a Hole, Secrets & Puzzles, Rags to Riches, etc.

Series Arc

Arc	Posts	Theme
Why Memory Matters	01–05	Stakes, taxonomy, pain
Why Existing Tools Fail	06–10	Competitive teardown
What Good Looks Like	11–15	Design principles
Build in Public → SilicaDB	16–20	Architecture + launch

Progress

• 01 — Your AI Agent Has No Memory (Dragon & The City)
• 02 — The Four Types of Memory Every AI Agent Needs
• 03 — Why Developers Keep Stuffing Everything Into the System Prompt
• 04 — The Agent That Learned the Same Lesson Twice
• 05 — Memory Is Now the Bottleneck. Not the Model.
• 06 — Redis Is Not Agent Memory. Neither Is a Vector DB.
• 07 — I Benchmarked Mem0, Zep, LangMem, and OpenBrain
• 08 — The Shared Blind Spot in Every Agent Memory Framework
• 09 — Why Temporal Reasoning in Mem0 and Zep Still Isn't Enough
• 10 — The Compliance Trap Nobody's Talking About
• 11 — Schema-Native Storage: What It Means and Why Agents Care
• 12 — Temporal Validity vs Temporal Reasoning
• 13 — What Would Agent Memory Look Like If We Designed It Like a Database?
• 14 — I Built Agent Memory on SQLite + FTS5 + sqlite-vec. Here's What Broke
• 15 — The Case for Embedded Agent Memory
• 16 — Designing SODL: A Schema Language Built for Agent Memory
• 17 — Why I Chose Zig (Then Reconsidered)
• 18 — The Four-Memory-Type Problem: How SilicaDB Models Agent Cognition
• 19 — SilicaDB + MCP: Making Agent Memory Accessible to Any Client
• 20 — Introducing SilicaDB

Posts committed one by one as research completes.