Memory of the person

The memory extension gives the agent a memory of the person it works for: who they are (their name, their role) and how they want things done (their corrections, how they like answers). It lasts across sessions and across all their projects, which a project's knowledge base can't: padawan learning how a student studies, course after course; miyagi learning a teacher's style.

Turning it on

import os from "node:os";
import path from "node:path";

const spec: AgentSpec<Config> = {
  // …
  extensions: ["memory"], // with others: ["sources", "knowledge", "memory"]
};

const config: Config = {
  // …
  memoryDir: path.join(os.homedir(), ".miyagi", "memory"),
};

• memoryDir is a folder of the agent's own, outside any project. Without it, the extension is left out and the agent can tell why.
• Never share it between agents. Different agents work for different people, or the same person in different roles, and what suits one spoils another. A shared memory would also be a way for an instruction planted in one agent to reach the others.
• One memory per agent and per user of the computer. If the agent works with several accounts and wants one memory for each, it puts the account in the path (~/.miyagi/memory/<account>).

What the agent sees

Its prompt section lists what it remembers, one line per entry, the most recently saved first (up to 50; past that, the line says how many are left out and recall gives them all):

## Your memory of the person
You keep a memory of the person you work for, across all their projects: …
- no-fish-puns (feedback): Dislikes puns about fish; avoid fish-themed jokes
- preferred-name-fran (user): Wants to be called Fran

Then its rules:

• Apply what it remembers without being asked.
• Save when the person corrects it, says how they like things done, or tells it something about themselves that will matter again, and say so in a few words ("I'll remember that"). It doesn't ask first. If an entry already covers it, update that entry instead of adding another.
• Keep only what's about the person wherever they work. How to do this project's work goes in the project (the knowledge base's preferences), and what only this task needs goes nowhere.
• Write the entries in the language it replies in.

The tools

Three, named like memory works (the model sees them as mcp__memory__recall and so on):

Tool	What it does
recall	With name, that entry whole; without it, every entry (name, type and description). Read-only.
remember	Creates an entry or changes one, with the person's quote. A new one takes name (kebab-case), type (user or feedback), description (one line) and body (for feedback, the rule, why and how to apply it). To change one, its name and only what changes: type, description or body whole, or old_string and new_string for a phrase of it, in its description or its body.
forget	Forgets an entry.

Changing an entry needs no recall first when the model knows what to change: a new description, say, which it sees in the prompt. old_string and new_string work like Claude Code's Edit (and have its names, which models know): old_string must be in the entry (description and body together) exactly once, or nothing changes and the tool says why (no line numbers, which shift and which the model would have to read first).

Only what the person said. remember checks that its quote appears in a message the person wrote in this conversation, ignoring case, spacing and quote marks. Otherwise it changes nothing, for a new entry or an existing one, and tells the model why. The extension hears the person's messages with a UserPromptSubmit hook (and, in a resumed run, reads the earlier ones from the run's conversation). A document, a web page or a tool result that says "remember that the person wants…" can't put anything in the memory, however it's worded.

The file tools never reach the folder (the denial points to its tools), and in plan mode only recall works.

Commands for the person

Command	What it does
/memory:list	Shows everything the agent remembers about them, by type.
/memory:forget <entry>	Forgets an entry, by name or description.

Saying it in plain words works too ("forget that I don't like puns"). With the awareness extension, the agent can tell the person the memory exists and how to see it.

On disk

One markdown file per entry, <memoryDir>/<name>.md, like Claude Code's auto-memory (which the kit keeps off, since it belongs to the runner and is per folder):

---
name: no-fish-puns
description: Dislikes puns about fish; avoid fish-themed jokes
type: feedback
updated: 2026-10-07T13:15:05.678Z
---

Rule: never tell jokes or puns about fish to this person. Why: they said "I can't stand puns about fish." Apply: …

Deleting a file forgets that entry; deleting the folder forgets everything.