IDEAS.md

Ideas parking lot

Good ideas that are out of scope right now are written here instead of built (build plan section 0: never expand scope mid-phase). The list is the owner's own, tidied on 2026-10-05: grouped, spelling fixed, and checked against the app.

Check the app before building anything here. Most of this list has been built since it was written, and the common mistake in this project is rebuilding something that exists (CLAUDE.md, section 1). The built items stay, with where they live, because their wording is what a future session will search for.

• Built
• Partly built
• Still open
• From the audits

Built

Capture and the shell

• Handle image and file uploads, with drag and drop: attachments, the Library's Images and Files views, captions, vision reads, Tesseract text and the OCR workspace.
• A guided setup on first install: the welcome card and the guided tour (Settings, Help), which walks through the real controls.
• A mini bar at the bottom with statuses and quick access to the command palette: the status bar at the foot of the window, and the palette on Ctrl/Cmd+K.
• Allow saving custom appearances and themes: thirteen themes and saved looks, in Settings, Appearance.
• Collapsible sidebars.
• Make MemoryMap AI cross-platform: a Windows installer and a Linux package, and macOS from source.
• Package the application, and improve the Settings models page: the installers, and Settings, Models with fit badges and a model per feature.
• Improve start.bat: no console window to keep open, and a tray icon. The launcher has --doctor, --logs and --shortcut, and the desktop window has a tray menu with Open, View logs, Restart and Quit.
• A way to exit the app and quit the backend: Quit in the header, and Settings, Background tasks.
• An interface for managing the application (backend, console, update, packages, restart, health): Settings, Background tasks, Packages, Logs and About.
• Improve the console in settings and show all console messages: Settings, Logs.
• A way to run the application on a phone: a layout built for the phone, the Share to MemoryMap action, and, for a phone on your own network, LAN mode over HTTPS (Settings, Account & security).
• The help area in settings has an ask-AI feature that knows the program's documentation: the Guide, and Ask Atlas about the app (Ctrl+Shift+H).
• A full security sweep: the audit of 2026-10-05 and its fixes are in the changelog under Security, and SECURITY.md states the model.
• Update the readme: rewritten on 2026-10-05, with every document in docs/.

The AI

• More tools, including managing categories (create, rename, merge, delete), and better agentic workflow: 65 tools, a visible plan, a confirm card for anything that changes or deletes.
• More skills, and a way to make new ones, including a full audit and clean-up of the notebook: 21 built-in skills (Notebook health check, Clean up my tags, Reorganise my categories, Fix my links, Find notes worth combining, Build a skill), and your own as Markdown files.
• Quick, normal and detailed modes for chat and agent: response presets.
• Reduce token usage in every AI interaction: the per-turn context budget, the small-model mode and schema compaction.
• Optimise outputs for what the current model can do: the small-model mode, and per-model sampling defaults.
• A nudge for the semantic search so "what did I save in the last two days" works: date-aware search operators and time travel over meaning.
• Reduce the cap on semantic search results, based on the model's context: search_notes scales its default result count with the model's window. The ceiling deliberately does not scale, because letting it do so meant a 128k model could pull 768 note previews into one tool result.
• Have the agent reachable from anywhere in the program: the popup agent (Ctrl+Shift+A) over every tab, and the agent activity monitor.
• Dynamically change models based on the task: background work uses the utility model so it does not tie up the chat model (smart_model_routing_enabled).
• Chat message metadata that survives a reload: the whole stats block is stored per turn, and every message says how full the model's window got.
• Make the notebook constellation regenerate on a light or dark change: it is rebuilt on every mode, accent and palette change, keeping the same arrangement.
• The background librarian: runs on an interval, tags, links and flags duplicates, off by default, never deletes. The category work (rename, merge, move) is done by the skills above.

Notes, documents and boards

• Expand and improve the sketches board, maybe a whiteboard tab: boards and mind maps in the Library's Boards & maps view, with shapes, connectors, an object library and mind maps.
• Better documents UI and usability, and make notes and documents two halves of a whole: the long-form editor, document links, backlinks and the Contents tree.
• A visual timeline: the Timeline tab with feed, table and scrubber.
• More dashboard widgets: the dashboard's widget set, including Boards & maps.

Partly built

• Agent asks for permission in chat and documents, with before-and-after comparison. The confirm card exists for destructive and outside-text-triggered changes. A before-and-after comparison for every edit does not.
• The agent controls the app on screen, with a visible border and a cancel. The popup agent and the activity monitor exist. The monitor is fixed bottom-right rather than movable or dockable, and the agent does not drive the interface.
• Dynamic model choice by task complexity. Routing is by caller, not by how hard the task is.
• Better agentic web search through chat. web_search and read_url exist, with a SearXNG option. A research mode that plans several searches does not.

Still open

• Manually group notes together, separate from the main sorting.
• A note with more than one category.
• Compress notes and data on disk.
• Gravity and spread have no effect on the graph's other layouts.
• A visual timeline drawn as a branching line with off-shoots.
• Settings pages cannot be reached on a narrow phone view (check for other pages with the same problem).
• No Markdown rendering in some of the dashboard's widgets.
• Pie charts for the dashboard (it has bar and line chart widgets).
• Different ways to sort chats, and grouping chats, with an agent tool and a skill to manage them.
• An agent permission flow shown as a dialogue in chat and documents, with the agent's edit proposed as a suggestion you can accept.
• An in-built browser with MCP tool abilities beside web search. (A stdio MCP server over the app's own tools exists, python -m memorymap.mcp_server; this is the other direction.)
• A VS Code extension.
• The status dot's "X" state in the top bar: it was never seen, and the AI status now lives in the status bar. Check that its error state can be reached.
• Streamline, enhance and optimise the backend and every AI interaction (ongoing).
• "I wrote 'ai is cool' as a note and it was filed under sketches": check that filing against the notebook's categories.
• Check docs/index.html's own text against the README, so the documentation site and the README say the same.

From the audits

• A dry run for the background librarian. It edits notes unattended and there is no way to see what it would do first. taskhistory already records each run; what is missing is "here is the diff, apply or discard".
• Sweep for orphans on a schedule, not just vectors. clean_orphaned_vectors is the right pattern, and clean_orphaned_board_cards follows it; attachments want it too.
• The sketch highlighter is at 5% opacity, which is roughly twenty passes before anything shows. Almost certainly a mistyped value, but it is a taste call, so it was left alone rather than changed during an audit.
• A memory-stream screen. Built: Settings, What it remembers lists everything the AI has saved, says how much of it reaches the model, and lets each one be edited, switched off or forgotten.
• Whiteboard cards should die with their note. Built: they are swept with the orphaned vectors on the background pass.