Agent Memory Kit: what to keep, what to forget, a memory file format and a checker for stale or conflicting facts

Run python3 mem_check.py check MEMORY.md, fix every FAIL, and list the WARN lines for me to decide.

From the pack.

# Agent Memory Kit: what your agent should remember, what it should forget, and a checker for both

A PaperChaseWebb pack, version 1.0, 2026-10-31. For any agent that keeps notes between sessions: Claude Code, Codex, Cursor, ChatGPT or Claude projects, Grok Bot, and your own agents. Plain Markdown plus one small Python script (3.8 or newer, nothing to install).

Agent memory goes wrong slowly. An old price stays "true" for a year, two time zones sit side by side and the agent picks one at random, a pasted key lives in a notes file, a line from a web page turns into a standing order. This kit gives you rules for what to keep and what to forget, one line format that carries its own source and review date, and a checker that flags stale, conflicting, duplicated and dangerous entries before your agent acts on them.

## What is in the kit

| Section | What it is for |
|---|---|
| 01 Keep or forget | What belongs in agent memory, what never does, and how long each kind should live |
| 02 The memory file format | One line per fact, with key, value, kind, source, dates and status, and a filled example file |
| 03 Changing your mind | How to update a fact without losing the history, and how to retire old ones |
| 04 The checker | `mem_check.py check`, `due` and `active`: every code and what to do about it |
| 05 A monthly review | Fifteen minutes a month that keeps memory true |
| 06 AGENTS.md | Rules an agent follows when it reads and writes memory |
| 07 How we tested it | Fixtures, unit tests, Python versions, and what was not tested |
| 08 The files | The script, its unit tests, and a good and a bad sample memory file |

## What you get

- **Keep or forget rules** for six kinds of memory (preference, fact, decision, project, person, temporary), with a lifetime for each.
- **A memory file format** that works in any Markdown file the agent already reads, and still reads well to a person.
- **A checker** that flags stale entries, passed review dates, two active values for one key, duplicates, stored keys or card numbers, sensitive personal data, lines that read like instructions, and missing sources.
- **A `due` list** of what to confirm in the next 30 days, and an **`active` view** that gives the agent only what is current.