Insights
Notes from building memory for AI agents.
Field write-ups on what actually works when you wire an AI coding agent to an external memory layer. Two formats: deep technical guides for the full picture, short problem-solution answers for a specific intent.
Deep guides
1 articleLong-form technical articles. The full thesis, the trade-offs, the honest limits. For builders who want to understand the architecture, not just the answer.
Problem-solution guides
5 answersShort, intent-driven answers for a specific question. Same operational substance as the deep guides, distilled to the action.
-
Problem-solution
Claude Code keeps forgetting your project? Here is what fixes it.
The short, operational answer to the Claude Code re-explanation tax: external memory with executable patterns, not a bigger context window.
-
Problem-solution
Cursor loses context on long projects. Here is the operational fix.
Cursor drifts on real codebases — re-asks about files it edited an hour ago, violates conventions across modules. The fix is memory scoped to the repo and file, wired through MCP.
-
Problem-solution
Operational memory for AI agents: what it is, what it is not.
Not a vector database, not a chat history, not a long system prompt. A typed, bounded, per-tenant catalogue of executable patterns the agent retrieves contextually at execution time.
-
Problem-solution
Persistent memory vs long system prompts: which one should you actually use?
Long prompts ship every turn. Persistent memory ships only what matches. The trade-offs, the failure modes of each, and when long prompts still win.
-
Problem-solution
How to make AI agent workflows reliable: executable patterns and stop conditions.
Reliability is a property of the loop, not the model. Executable patterns the agent applies, and explicit stop conditions that bound where the agent can act on its own.
Deep guide = 3000+ words, full architecture and trade-offs. Problem-solution = short answer to a specific search intent. Both share the same operational substance.