Petty-Memory

What you get

• Capture from anywhere. The desktop app (quick-capture window, drag and drop), the browser extension (popup, side panel, right-click, one-click import of your browser bookmarks, offline queue), the iOS app and its share sheet, the CLI, and any AI tool over MCP.
• Pages that outlive the web. Saved links are fetched and stored as readable Markdown, and PDFs and text files are extracted, so everything stays searchable even after the original page disappears.
• Search by words and by meaning. Full-text search plus a local embedding model (no cloud), fused so both exact terms and paraphrases surface, with a relevance floor so irrelevant results don't crowd in.
• Project memory for AI agents. Each project has standing instructions, a living memory document (every version kept), an append-only decision log, tasks, and a briefing an agent reads in one call.
• Formatted editing. Shared rich-text and Markdown editing, note split/join, and undo/redo for text and structural changes.
• Ordered outlines. Nest items inside other items, keep their sibling order, and move a parent with its children. The desktop Outline and the CLI, MCP and HTTP API use the same core rules; existing notes and project filing stay intact.
• Organize, review and recover. Headings and section breaks, attention levels, review marks, highlights and nested checklists; inline item references with backlinks; per-item history with compare and restore; and optional AI edit proposals you review, apply selectively and can undo.
• Secret notes and public links. Passphrase-encrypted notes, nested secret branches and encrypted files, sealed on the device in the browser and on iPhone, with portable recovery kits. Read-only public links to a chosen subtree, with parts left out, expiry and revocation. See the security model; these features have engineering and interoperability tests, but no production security audit.
• In the browser. A browser client served by the desktop app or a small self-hosted launcher (apps/web), for your own devices.
• Shelves and reminders. Pinned, Read, Watch, Buy, Revisit, "unopened for a month". Reminders take plain words ("tomorrow", "friday") and notify on the desktop, the phone or in the browser.
• Saved views. Name a recurring search such as "Open Garage tasks" or "Security articles to read". Combine keywords, project, type, tag, task status and attention filters; the results follow the current library. Available on desktop, web, iPhone, CLI and MCP.
• Reusable templates. Turn an ordinary note/checklist tree into an editable template, then create fresh copies for meeting notes, research briefs or recurring checklists. Copies keep nesting and references, reset completed boxes and expand {{date}} and {{project}}. Examples and limits.
• Chat with your library. Ask questions with source links, discuss follow-ups, or say “organize all notes about gardening into a PARA system.” Review the proposed hierarchy and note placements, keep the branches you want, apply them together and undo safely. Ask for tasks, tags, reminders or a brief from a branch or topic, and review each before applying. Chats are kept only when you save them. Dictate into the draft on iPhone and supporting browsers, with system dictation available on desktop. Chat and voice guide.
• Review and connect what you know. A daily review queue brings notes back on a schedule you choose. Find connections suggests unlinked mentions, related notes and possible duplicates, and duplicates can be merged with a preview and undo. Neither needs AI. Everyday workflows.
• Text from images, highlights from books. Extract text from an image locally and review it before saving, with the image kept. Import Enve Book Player highlights and annotations; importing again skips what you already have.
• Optional AI, bring your own. Ollama, OpenAI, Anthropic, Gemini, OpenRouter or any OpenAI-compatible server for summaries, tag and project suggestions (or automatic filing), and cited answers from your own library. Everything else works with AI off.
• Automations. For example, "links from github.com → #code, file into Development".
• Sync between your computers through a folder you already sync (iCloud Drive, Dropbox, Syncthing), with optional passphrase encryption, or through a Petty Memory server you run yourself. Server sync watches local changes and long-polls for incoming work; folder sync runs periodically. Conflicting text is retained for review.
• Petty appearance. Cool gray surfaces, teal accents, light/dark themes and true-black OLED on iPhone.
• Yours to keep. Automatic hourly, daily and monthly backups; Export Everything to Markdown, JSON and the original files; importers for browser bookmarks, Markdown folders or Obsidian vaults, bookmark CSVs, and Enve exports.

Daftak feature comparison

This work began as a public challenge to build an open-source alternative to Daftak. Petty Memory is an independent implementation under AGPL-3.0-only. The table maps all 17 feature groups on Daftak's public page, reviewed October 17, 2026, to implemented Memory capabilities.

Advertised capability	What Petty Memory implements
Nested items	Ordered, paged outlines; move, indent and outdent whole branches while keeping stable item identities.
Automatic formatting	A formatted editor with toolbar controls, semantic headings and checklists, plus optional Markdown source. Existing Markdown stays intact.
Automatic organization	Opt-in background organization, selected-branch proposals and natural-language topic organization across the library, with review, selective acceptance and guarded undo.
Attention levels	Raise or lower an item's visual emphasis without changing its content or filing.
Checklists	Nested completion and progress, completed-item visibility and quick reset.
Connections	Create and remove relations between items; browse related items, backlinks and the desktop graph. Suggested connections are offered without AI.
Item history	Automatic item/subtree snapshots, comparisons and restore as a new change; historical attachment bytes are retained.
AI assistance	Chat with optional saving, dictated prompts, cited answers, scoped editing, suggested actions and topic-wide grouping with selective acceptance and guarded undo. Bring your own provider or keep AI off.
Inline references	Insert references to stable item identities from the editor and navigate to the referenced item.
Hybrid encryption / secrets	Optional encrypted notes, nested secret branches and encrypted attachments, with locking, recovery kits and scoped device access. See the encryption differences below.
Selective sharing	Read-only public subtree links with excluded branches, expiry and revocation. Secrets are excluded.
Highlights	A highlights view, explicit highlighting/pinning and optional AI suggestions.
Tags	Tag capture, editing, filtering and search across the library.
Section breaks	Labeled dividers between unrelated sections of an outline.
Review marks	Mark items that need improvement or organization and find them again.
Instant sync	Local change watching and server long-polls; foreground browser/iPhone refresh; optional encrypted folder sync runs periodically.
Data ownership	Import, export, deletion, portable Markdown/JSON/original files and full-library backups.

All 17 groups have implementations. The verification record lists the tested workflows and remaining release checks; this is not a claim of identical signed-in behavior or production certification.

A few differences matter when choosing how to run it:

• You host the library. Use the desktop app or the self-hosted browser/server setup. A managed account service has not been deployed, and the local product needs no account.
• Encryption is selective. Secret payloads and files are encrypted; ordinary SQLite content and some structural metadata remain readable on the host. Configure TLS for remote access. Security details explain key handling, recovery and limits.
• iPhone keeps an ordinary offline library. Browse, search and queue edits offline. Secret notes, server history, public links, AI, saved-chat resuming, daily review, connection suggestions and imports need a connection. Saved transcripts remain readable as ordinary cached notes; attachment listings are cached, file bytes are not mirrored automatically.
• Release work remains. Physical-device/background/share-sheet acceptance, full accessibility/IME and performance checks, Windows/Linux installer validation, signing and store distribution still need their respective environments. Daftak lists native mobile apps, collaboration and image support as roadmap work; Memory already has iPhone and image capture, while team collaboration is outside this comparison.

Workflows beyond the comparison

The comparison above covers Daftak's 17 public feature groups. These additions go further, and most work without an AI provider:

• Saved chats. Opt in per conversation. A saved chat is an ordinary plaintext note that syncs, exports and searches like any other. Unsaved chats are not stored.
• Suggested actions. Tasks, tags, reminders and a brief from a branch or topic. Each operation is previewed. Selected operations apply together, stale proposals are refused, and Undo stops before overwriting newer work.
• Broader topic organization. Keyword discovery stays complete. Up to 50 related-in-meaning notes are added and marked Suggested, with keyword-only results when semantic search is unavailable.
• Voice and image capture. Dictate a note, or extract text from an image and correct it before saving. The original image is kept. Image text is read locally: by Tesseract (English) on the library host for desktop and browser, and by Vision on iPhone.
• Daily review. A local schedule and review history on each item. Choose when to see each item again. No background service or extra notifications.
• Connections and duplicate cleanup. Unlinked mentions, related notes and possible duplicates are found without AI. Merges are previewed, keep the original note archived with its files, and can be undone.
• Enve book highlights. Preview and import highlights, annotations and bookmarks from Enve Book Player's JSON or Markdown export. Importing again doesn't create duplicates. Enve's export has no book identifiers or source links, so there is no live sync or deep link back into a book.
• Getting started. A first-run checklist on desktop, browser and iPhone. AI and accounts stay optional.

The Knowledge workspace adds eight further workflows:

• Reusable AI commands. Save a named instruction and scope, then run it against fresh notes. Review tasks, tags, reminders and briefs before Apply; guarded Undo remains available.
• Explicit organization preferences. Save filing and organization instructions, optionally per project, and choose which ones each workflow or organization request should use.
• Typed collections. Add text, number, date, checkbox and select fields, then use table, card or Kanban layouts with typed filtering and sorting.
• Audio research. Record or import audio, preview/save timestamped captions, seek to cues and keep source-linked passages. Optional transcription uses a separately configured speech server with confirmation for each recording.
• Biscotti meetings. Preview and import selected meeting summaries, notes and speaker transcripts on desktop, in the browser or on iPhone. The Mac desktop app can browse Biscotti directly and open recording timestamps. Prepare for the next meeting, draft follow-ups or a brief, and schedule a review through the existing AI and review tools. Reimports preserve your edits. Meeting guide · Fictional sample CSV.
• PDF research. Read and search by page, annotate exact passages, recognize scanned pages into reviewed text, and follow AI citations to a page or recording timestamp. Originals are retained.
• Watched sources. Opt into checks for pages, RSS/Atom feeds or a locally approved folder. New and changed material waits for selected import or update, preserving originals and history.
• Information review. Find older notes, mark them checked and review possible conflicts using exact quotes and separate dates. Explicitly choose which ordinary note supersedes another; AI never silently rewrites claims.

Knowledge workflows explains these features, limits and privacy boundaries. Verification records their actual acceptance checks.

Typed collections	Reviewed AI workflows

Everyday workflows explains each workflow, its limits and its privacy boundaries. Verification records the checks for this batch; physical-device acceptance for microphone, camera and Vision text recognition is still outstanding.

Saved views and templates came first. For example, create a view named Garage work with project Garage Door, type Task and status Open. It updates as tasks are added and completed. Save a Weekly review note with nested checklist entries as a template, then choose a destination project when creating the next review. Both workflows work without an AI provider.

Desktop and browser evaluate views locally on their library host. iPhone can display cached lists/results when disconnected; creating or editing views and copying template trees requires the paired server. These are keyword/filter views, with cursor paging, rather than saved semantic rankings. Workflow guide includes CLI and MCP usage and the research references behind this batch.

Saved view	Note created from a template

Chat and organize notes

Configure your provider in desktop Settings → AI, or with petty-memory ai use on a self-hosted server. Your phone and browser use that host's provider.

In AI chat → Organize library (Organize on iPhone), type or dictate:

Organize all notes about gardening into a PARA system.

Memory finds matching notes across projects and proposes a destination, headings and placements. Review the tree, untick anything that does not belong, then move the selection together. Note contents, files, identities and project filing stay intact. Undo protects newer work. Questions and follow-ups have source links; Change notes proposes edits to a branch you choose. The guide explains keyword matching and suggestions, size limits, saving chats and provider privacy.

Dictation fills the draft and never sends or applies automatically. iPhone uses native speech recognition; supporting browsers use Web Speech. Desktop offers system dictation guidance when in-app speech is unavailable. Live microphone acceptance still needs a physical-device run.

Choose Suggest actions (Actions on iPhone) for tasks, tags, reminders or a brief. Each operation is reviewed before applying; the examples below use fictional notes.

Desktop action review	iPhone action review

Desktop organization review	iPhone organization review

Screenshots

Outlines and formatted editing

Desktop outline	Browser editor

Desktop

Home. Quick capture, reminders, recent items.	Project briefing. Instructions, a living memory document, tasks: what an AI reads first.
Decision log. Append-only; replaced decisions stay visible as history.	Archived pages. The readable text is saved locally, next to your note, tags and reminder.
Search by meaning. "How does the remote avoid replay attacks" finds the rolling-code article.	Inbox with AI suggestions. Summaries, tags and a project to accept, or file automatically.
Activity. Which AI did what: Claude Code, Codex, the extension, your phone.	Graph. Items, projects and tags, connected.
Connect AI tools. One-click setup for Claude Code, Codex and JSON-configured clients.	Devices. Scoped tokens and QR pairing for your phone.
Ask your library. Answers only from what you saved, with sources; replaced decisions give way to current ones.	Light theme, or follow the system.

Browser extension

Side panel. This page, related items from your library, and shelves.	Popup. Save with project, tags, intent, reminder and the selected text.	One-click import of your browser's bookmarks, folders becoming tags.

iPhone

Home keeps quick capture, Ask AI, Library and recent notes in view. Open Library for the outline, daily review, shelves, saved views, templates and other tools.

Home	Outline	Formatted editor

Screenshots use disposable fictional libraries. The verification record includes light/dark captures and the tested recovery workflows.

Install

Build from main for the feature layer described here; published installers follow their own release version.

• Desktop (macOS): download the .dmg for Apple Silicon or Intel from Releases. The builds aren't notarized yet, so the first time, right-click the app and choose Open. Intel Macs use keyword search: the local embedding runtime has no Intel build.
• Windows and Linux: installers are configured (NSIS, AppImage, deb). Build them with npm --prefix apps/desktop run dist:win or dist:linux on that OS.
• Browser extension: run node apps/extension/scripts/build.mjs, then load apps/extension/dist/chrome in chrome://extensions (Developer mode → Load unpacked). Firefox: load dist/firefox from about:debugging. Create a token in the desktop app under Settings → Devices.
• iPhone: after npm install, run sh apps/ios/scripts/build-editor.sh, then generate the project in apps/ios with XcodeGen, set your team in Xcode, and run. Pair by scanning the QR code in Settings → Devices. See the iPhone build guide.
• CLI and MCP server only: Node 24+, then npm install in this repository (see Try it below).

How it works

flowchart LR
  subgraph You
    D[Desktop app]
    E[Browser extension]
    P[iPhone + share sheet]
    W[Browser client]
    C[CLI]
  end
  subgraph AI["AI tools"]
    CC[Claude Code]
    CX[Codex]
    O[ChatGPT, Cursor, …]
  end
  D --> Core
  C --> Core
  E -- "local API · scoped token" --> Core
  P -- "HTTP API over LAN / Tailscale" --> Core
  W -- "scoped HTTP API" --> Core
  CC -- "MCP (stdio)" --> Core
  CX -- "MCP (stdio)" --> Core
  O -- "MCP (HTTP)" --> Core
  Core["Petty Memory core<br/>services · rules · workers"] --> DB[("memory.sqlite<br/>FTS5 + vectors")]
  Core --> Files[("attachments/<br/>backups/")]
  Core -. optional .-> LLM["Your AI provider<br/>(Ollama, OpenAI, Anthropic…)"]
  Core -. optional .-> Sync[("Shared folder<br/>encrypted sync")]
  Core -. optional .-> Remote["Self-hosted server<br/>live sync + public links"]

Every client goes through the same core services, so the rules hold everywhere:

• every change is logged, with the client that made it
• decisions are append-only
• nothing an AI does can delete your data
• saved content is always treated as data, never as instructions

Try it

Requires Node 24+.

npm install
npm run cli -- project new "Garage Door" --instructions "Must work without internet."
npm run cli -- link https://example.com/security-plus --title "Security+ 2.0 protocol" -p garage
npm run cli -- note "Need a bench rig before touching the real opener" -p garage
npm run cli -- decide garage "Use ESP32-S3" --reason "USB host support"
npm run cli -- project show garage
npm run cli -- search opener protocol

The desktop app is in apps/desktop: npm --prefix apps/desktop run dev, or run dist:mac for the installers. The browser extension is in apps/extension, the iOS app in apps/ios and the browser client in apps/web (npm --prefix apps/web start -- --home ~/em-demo, then connect with a token from npm run cli -- --home ~/em-demo clients add "Browser" --scope read,write); each README explains setup.

The planned Daftak feature layer is implemented and has passed the documented functional checks. Release and external validation gates remain explicit.

Connect your AI

npm run cli -- connect

This prints ready-to-paste setup for Claude Code, Codex and JSON-configured clients such as Claude Desktop and Cursor. The desktop app has the same setup under Settings → AI tools, with one-click "Add to Claude Code". Then ask your assistant "What do we know about my garage door project?"

The MCP server exposes search, briefings, notes, links, files, tasks, ordered outlines, presentation and checklists, item history, references, project memory, decisions, reminders and shelves. None of its tools can delete anything, decrypt secret notes, publish shares or replicate a library, and saved content is always presented to the model as data, never instructions. See docs/MCP.md. Devices and remote agents use the local HTTP API with scoped tokens.

Where your data lives

OS	Folder
macOS	~/Library/Application Support/Enve Memory/
Windows	%APPDATA%\Enve Memory\
Linux	~/.config/Enve Memory/

Override it with --home DIR or ENVE_MEMORY_HOME.

Docs

Product · AI chat · Everyday workflows · Saved views and templates · Architecture · Data model · MCP · HTTP API · Security · Sync · Roadmap · Parity plan · Verification · Self-hosting · Contributing

License

AGPL-3.0-only. Plugins and clients that talk to Petty Memory over MCP or its API are separate programs.