Session log protocol for working with Claude

A chat with an AI agent has no memory worth the name. Close the window and the next session starts from a blank line. The fix is not a bigger context window. It is one row per session, written at the end, read at the start. The agent does the typing.

This is the protocol as it runs in my own Notion workspace, as an 8-page PDF.

What is inside

• The start ritual: read the last three rows for this area and the last seven digests across all areas, then answer in exactly one line. Why one line.
• The close ritual: the trigger word, the propose-confirm-write rule, and every field of the row: date, area, project, type, summary, key outcome, decisions, next, and a digest of 400 characters or fewer.
• The two body blocks, ## Next and ## Bugs / Issues, and the rule for when a bug becomes a task.
• The decisions log: which decisions earn their own row, and why the Why field is the whole point.
• A Notion schema to copy: two databases, every property with its type and options, and why you want filtered views instead of a second log database.
• Four prompts to paste into a chat project: start, close, mid-session decision, digest rewrite.
• Eleven failure modes I have hit, and what the protocol does about each.

Who it is for

Anyone running more than one track with an AI agent and tired of re-explaining the project every morning. Notion is the reference, but the same two tables work as markdown files.