memTools desktop companion

memTools is a Tauri 2 desktop quick-capture app for the local rsrs runtime. It is separate from the desktop client: the client manages a runtime and account, while memTools is a hotkey panel for writing and reviewing memories.

• Repository: https://github.com/risense-ai/respire-memtool
• Connection: MCP over HTTP (JSON-RPC 2.0) to http://127.0.0.1:15169/mcp
• Hotkey: Ctrl+Alt+R toggles the panel; Esc and the close button only hide it to the tray
• Window: drag any edge or corner to resize (minimum 360×400); the panel remembers its size and position. Narrow windows move secondary graph controls into the Display menu and trim the legend; wide windows split the capture view into content and destination columns

What it does

Area	Behavior
Capture	Main area or diary, one of the 23 taxonomy roots, then a type (context / decision / preference / task / emotion / time / skill)
Duplicates	On a suspected duplicate it offers change / merge / attach / store-new, and backs up before destructive choices
Automation	Local rules for tier and tree, two recall rounds for duplicates, optional OpenAI-compatible endpoint for tier/tree; falls back to local rules with an explicit warning
Diary	Written as importance=trivial, skipping duplicate checks and tree attachment
Review	Recent memories, keyword search, semantic rsrs recall, reminders for 【触发】 dates, read-only 3D graph from onememory.db
Capture surfaces	System tray, browser extension, Windows selection bubble, Explorer "add to knowledge base"
Knowledge base	Files, folders, a page URL, or selected text go through the AI window, Explorer, the selection bubble, or the browser extension; each item becomes a folder with original/ and <title>.md
Updating	Reads update.json from GitHub Releases, verifies SHA-256, self-replaces on Windows

memtools.exe --ai opens the AI window directly.

Install

Download from GitHub Releases: memtools-<version>.exe (portable), .msi, or the NSIS .exe. Release builds are unsigned and not notarized, so Windows warns about an unknown publisher. On macOS, if Gatekeeper refuses the app, right-click it in Finder and choose Open, or run xattr -dr com.apple.quarantine /Applications/memTools.app; System Settings → Privacy & Security → Open Anyway also works. From 0.8.1 the macOS DMG is universal (Apple Silicon and Intel); older releases ship aarch64 only, so Intel Macs need a newer build.

On first start memTools checks for the rsrs command and installs the CLI through npm or pnpm when it is missing. The runtime itself still has to be running on port 15169; memTools never stops or restarts it unless you confirm the restart in the rsrs panel.

Configuration is read in this order:

1. MEMTOOLS_CONFIG environment variable
2. %APPDATA%\com.memtools.desktop\config.toml (macOS ~/Library/Application Support/..., Linux $XDG_CONFIG_HOME/...)
3. config.toml next to the executable (portable)
4. Built-in defaults

AI window and knowledge base

The AI window is a separate window (top-bar AI button) that both chats with a local agent CLI and ingests material into a knowledge base.

Layout of one ingested item:

<knowledge base root>/<year>/<month>/<topic>-<timestamp>/
  original/     source file or directory, unchanged
  <title>.md    organized Markdown

The agent writes the result into rsrs following the change / merge / attach / store-new rules, and records the file path plus device in the body.

Ingest material

1. Open the AI window from the top-bar AI button or with memtools.exe --ai. Closing the window only hides it; the button or --ai brings it back.
2. Choose the Knowledge base directory on first use. The choice is stored in knowledge_base.json and overrides config.toml.
3. Add material in any of these ways:
   • Add folder, or drag files and folders onto the window.
   • Paste a path into the path field, or a page URL into the URL field.
   • Explorer right-click → Add to memTools knowledge base (Windows only).
   • Selection bubble → AI 调研, or the browser extension's right-click items; both enqueue through POST /kb/add.
4. Pick the agent and model (below), then follow the task list. The agent copies the original, writes <title>.md, keeps research notes for URLs and selected text, and stores a memory with the change / merge / attach / store-new rules.

Limits: only text-like files are parsed. PDF, Word, and other binary files are copied into original/ and registered, but their body is not extracted; convert them to text first. Defaults: 20 MB per file, 50 files per task, 600 seconds per task.

Agent sources

The source dropdown lists five entries and probes each one live:

Source	Needed on the machine	Detection
opencode	npm i -g opencode-ai	opencode --version; models from opencode models
Qoder CLI	Qoder IDE (bundles qodercli)	qodercli --version, then qodercli status for the account line
Claude Code	npm install -g @anthropic-ai/claude-code	claude --version; runs as a custom agent
Codex CLI	npm install -g @openai/codex	codex --version; runs as a custom agent
Custom CLI	Any CLI that takes a prompt and prints the answer	--version, or accepted as-is when the configured binary exists

When a source is missing or not logged in, selecting it expands an install and configure panel with the reason, the install command, the login steps, the parameters that will be used, the detected path, and a documentation link. The command can be copied with one click.

After installing, press Detect again, then Bind. Binding writes the choice to knowledge_base.json, which overrides config.toml.

Qoder's qodercli ships inside the IDE and is not on PATH, so a bare qodercli in PowerShell fails with "term is not recognized". Use the absolute path, or add the directory to PATH:

& "D:\...\Qoder\resources\app\resources\bin\x86_64_windows\qodercli.exe"
# then type /login inside the interactive session

The login hint in the window carries the detected absolute path and the & "…" call syntax.

Models

The model field is a free-text input backed by a suggestion list:

Source	Model values
opencode	provider/model, for example opencode/big-pickle; the list comes from opencode models
Qoder CLI	auto, efficient, gmodel, kmodel, lite, mmodel, performance, q35model, qmodel, ultimate
Claude / Codex / custom	Whatever the CLI accepts; empty means the CLI default

Switching sources clears the field, because the model vocabularies do not overlap.

Driving a custom CLI

Claude Code and Codex CLI are bound as kind = "custom" with prefilled parameters. The same mechanism covers any other CLI:

[processor.agent]
kind = "custom"
bin = "crush"                                    # command name or absolute path
args = ["run", "--model", "{model}", "{prompt}"]  # one array element per argv
env = ["CI=1"]                                    # KEY=VALUE
models = ["claude-sonnet", "gpt-5"]               # optional model suggestions
output = "json"                                   # text | json | opencode-json | qoder-json

Placeholders: {prompt}, {model}, {session}, {endpoint}, {workdir}, {agent}. An element that becomes empty after substitution is dropped instead of being passed as an empty argument.

Output parsers:

output	Expected stdout
text	Plain text; the whole stdout becomes the answer
json	One JSON object; reads result, text, content, response, output, then message
opencode-json	NDJSON event stream; takes the text of the last message
qoder-json	{"type":"result","result":"…"}

The binding keeps the interface choice above config.toml: knowledge_base.json stores the source, binary, model, and — for custom sources — the argument template, environment, and output parser. Clearing a field falls back to config.toml.

Knowledge base settings

The AI window stores its choices in knowledge_base.json, on top of config.toml:

Key	Default	Meaning
knowledge_base.enabled	true	Set false to reject POST /kb/add
knowledge_base.root	empty	Root directory; first-use picker
knowledge_base.agent_kind	opencode	opencode / qoder / custom
knowledge_base.agent_model	empty	provider/model for this window
knowledge_base.max_file_mb	20	Per-file limit in MB
knowledge_base.max_files	50	Files per task
knowledge_base.timeout_secs	600	Agent timeout in seconds
knowledge_base.context_menu	true	Explorer right-click menu (Windows)
knowledge_base.research_websearch	true	Web research for URLs and selected text

Platform notes

Capability	Windows	macOS / Linux
Capture, review, search, reminders, graph, tray, hotkey	Yes	Yes
AI window and knowledge base ingest	Yes	Yes (native folder picker is Windows-only; type the path elsewhere)
Explorer right-click "add to knowledge base"	Yes	No
Selection bubble	Yes	No
Automatic update download and replace	Yes	Check only

The browser extension talks to the local endpoint on port 17650; its token lives in extension_token.txt in the app config directory.

Troubleshooting

Symptom	Cause and fix
runtime 不可达	rsrs runtime is not listening on 15169; start it with rsrs status or the host application
A source shows not ready	Open that entry and follow the install panel: install, log in, Detect again
xxx is not recognized as a cmdlet	The CLI is not on PATH; use the absolute path from the panel, or add its directory to PATH and reopen the terminal
Binding succeeds but the first run fails	Re-check the version with Detect again, and confirm the model format (provider/model for opencode)
config.toml edits seem ignored	The AI window choice in knowledge_base.json wins; bind again after editing the file
Windows warns about an unknown publisher	Release builds are not code-signed

See also

• CLI for the memory commands the agent uses
• Runtime RPC for the local protocol
• Client for the separate Wails desktop client