DeskLore

Your computer's memory. Owned by you, usable by the agents you trust.

Open-source personal context infrastructure for macOS. Local. Cited. Deletable. Yours.

简体中文

_{Real DeskLore interface with a fully synthetic activity history.}

DeskLore is open-source infrastructure for personal context. It observes what you do on your Mac through macOS Accessibility semantics instead of screen recording, and turns that activity into a memory stored on your machine as plain files. Every statement in that memory can be traced to the evidence behind it and deleted together with it. DeskLore uses this memory itself first; agents you trust can use it within the scope you grant.

Today DeskLore ships the evidence layer: a searchable, multi-resolution timeline with citation-validated summaries. The memory layer and its outputs, including proactive assistance and agent access, are the direction described in docs/DIRECTION.md.

Status: 0.2.0 is an early source release for macOS 14+ on Apple Silicon. There is no official signed binary yet.

Why DeskLore

• Memory is the product; history is its evidence. Timeline documents at ten-minute, six-hour, and daily resolutions are the retained evidence that every later claim must cite.
• Semantic capture first. App, window, interaction, URL, and Accessibility context are normalized into evidence-aware events. Screen capture is a fallback and stays off by default.
• Local-first storage. Raw events, Markdown timelines, timeline rollups, settings, and encrypted API credentials stay on your Mac unless you explicitly enable a model-backed feature.
• Explicit consent per boundary. The native collector does not start before the first-run consent screen is accepted, and each capability that sends data off the machine is a separate decision.
• Built as infrastructure. Evidence producers, the memory layer, and consumers are separate layers with stable, versioned formats, so new collectors and agents can be added without rewriting the core.
• Bilingual interface. English is the default; Simplified Chinese can be selected during onboarding or later in Settings.
• Inspectable artifacts. Timeline details and rollups are Markdown files with source IDs and deterministic fallbacks.

Privacy defaults

After explicit consent, DeskLore observes ordinary applications and URLs by default. It excludes DeskLore itself, sensitive macOS system surfaces, private-browsing windows, and password-like fields. Recording can be paused at any time.

Capability	Default	Network	Retention
Semantic events	On after consent	None	Raw segments: 48 hours
Timeline and rollups	On	None with deterministic summaries	Until deleted
Visual fallback	Off	None for capture; optional model use	Text evidence: 24 hours; pixels are not stored
Model summaries	Off	User-configured HTTPS endpoint	Generated Markdown stays local
Telemetry	Not included	None	N/A

Deleting a timeline item also deletes its source segment and related visual evidence; affected timeline rollups are regenerated without it. Clear all history removes raw events, timelines, rollups, and visual evidence, then leaves recording paused. See PRIVACY.md for the complete boundary.

Requirements

• macOS 14 or newer
• Apple Silicon
• Node.js 24+
• pnpm 11+
• Xcode Command Line Tools with Swift 6.2+

Run from source

git clone https://github.com/FoundDream/desklore.git
cd desklore
pnpm install --frozen-lockfile
pnpm dev

On first launch, DeskLore remains stopped until you accept the local recording boundary. macOS then asks for Accessibility permission for DeskLore Collector, the embedded native capture component. Screen Recording permission is requested only if you later enable visual fallback.

Useful commands:

pnpm check           # format, lint, and TypeScript checks
pnpm test            # TypeScript, evaluator, and Swift tests
pnpm native:test     # Swift collector/core tests only
pnpm build           # production Electron and Swift build
pnpm package:mac     # local DMG and ZIP; signing depends on your keychain

Architecture

React renderer
  -> narrow preload API and validated IPC
  -> Electron main process
     -> ServerCore utility process
        -> policy, coalescing, storage, timeline, rollups, optional model calls
        -> NDJSON over stdio
           -> DeskLore Collector (Swift)
              -> Accessibility, AXObserver, NSWorkspace, global interaction events
              -> native redaction, optional ScreenCaptureKit fallback

The renderer receives sanitized DTOs only. API keys and raw JSONL never enter the renderer. Collector and UI have separate bundle identifiers so macOS can sign and authorize the native capture boundary independently.

See docs/ARCHITECTURE.md for the source layout, dependency rules, process boundary, and data ownership model, and docs/DIRECTION.md for the layer model and the current phase.

Local data lives at:

~/Library/Application Support/DeskLore/history/
  segments/       # 10-minute raw JSONL buckets, retained for 48 hours
  timeline/       # derived Markdown timeline cards
  rollups/6h/     # six-hour Markdown rollups
  rollups/day/    # daily Markdown rollups
  state/          # consent, language, policy, visual, and model settings

Files are created with owner-only permissions where supported. DeskLore does not add application- level encryption to timeline or rollup files; use macOS FileVault if disk-at-rest protection is required.

DeskLore's public release line uses a new, versioned local schema. It does not import data or unversioned settings created by pre-release Computer History builds.

Optional model features

DeskLore works without an API key. Deterministic rules create timeline details and rollups offline. If you enable model summaries, DeskLore sends filtered evidence to the HTTPS endpoint you configure. API keys are encrypted with Electron safeStorage and are not exposed to the renderer. Model settings support both OpenAI Responses and Chat Completions wire protocols, including compatible custom endpoints.

Visual fallback has three independently configured parts:

1. AX sufficiency judgment (rules by default, optional model judgment).
2. Window capture (off by default).
3. Understanding (off, local OCR, or an explicitly configured model).

Raw screenshots are processed transiently and are not written to event files. See docs/EVALUATION.md before making quality claims from the included benchmarks.

Contributing

Read CONTRIBUTING.md before opening a pull request. Privacy, deletion semantics, native permission boundaries, and evidence integrity are compatibility contracts in this project. Security issues should follow SECURITY.md.

License

Copyright 2026 Ziwen Song.

Licensed under the Apache License 2.0.