Blog
Notes on memory for AI agents.
Why AI coding agents keep making the same mistakes
An agent re-proposing an approach you ruled out last month isn't reasoning badly. It's reasoning correctly from information that no longer includes the failure.
Read the post →Claude Code context compaction: what gets lost, and how to preserve the knowledge that matters
Compaction keeps a long session alive by summarizing its own history. Summarization drops specifics first, which is exactly the part that took an hour to earn.
Read the post →How to make Claude Code remember project context across sessions
Five mechanisms, one working combination. Session boundaries, CLAUDE.md, built-in memory, context windows and MCP: what each actually covers, and how to layer them.
Claude Code keeps forgetting your project? How to add persistent memory
It isn't a model failure, it's a storage failure. What Claude Code keeps between sessions, what it drops, and how to add persistent project memory in one command.
Read the post →CLAUDE.md vs persistent project memory: when you need both
One is loaded in full every turn; the other is searched on demand. That single difference decides what belongs in each, and why a growing CLAUDE.md is a symptom, not a failure of discipline.
MCP memory servers explained: persistent memory for AI agents
A memory server is the MCP server that gives an agent somewhere to put what it learns. Here's the category, the five designs in it, and the one distinction that decides whether you can trust the results.
Read the post →How to share project context between Claude Code and Codex
Different instruction files, different sessions, no common store. Here's how to give both agents one shared project memory so whatever either learns, both can use.
Read the post →How to give Claude Code and Codex persistent memory over MCP
A step-by-step walkthrough: connect any MCP client to a shared memory server, then have your agents store decisions with upsert_note and recall them with search_notes.
Agent handoffs: keeping context when one session ends and the next begins
Long tasks outlive a single context window. Here's how to write a handoff an agent can actually pick up: what to capture, what to leave in git, and how to structure it.
Read the post →Why AI agents need persistent memory (and how to give them some)
Agents reason brilliantly in the moment and forget everything the moment the session closes. Why that happens, why bigger context windows don't fix it, and what a real memory layer looks like.
Read the post →What is the Model Context Protocol (MCP)? A plain-English guide
MCP is the "USB-C port" for AI tools, one standard way for agents to reach external data and actions. What it is, why it matters, and how memory fits in.
Read the post →