dachshund

agent
Guvenlik Denetimi
Basarisiz
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 9 GitHub stars
Code Basarisiz
  • process.env — Environment variable access in scripts/build-mcp.mjs
  • network request — Outbound network request in scripts/fetch-agent-icons.mjs
  • process.env — Environment variable access in scripts/screenshot.mjs
  • fs.rmSync — Destructive file system operation in scripts/seed-demo.mjs
  • spawnSync — Synchronous process spawning in scripts/test-agents.mjs
  • fs.rmSync — Destructive file system operation in scripts/test-agents.mjs
  • exec() — Shell command execution in scripts/test-agents.ts
  • fs.rmSync — Destructive file system operation in scripts/test-agents.ts
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

One memory across all agents.

README.md

dachshund

dachshund

Every agent's memory. One local place.

Claude Code remembers one decision. Codex remembers another. Cursor remembers the exception.

dachshund finds the memory files your agents already keep, brings them into one
Markdown library, and serves that library back through MCP. No account, no cloud,
no model requirement.

Download for macOS Status License Local

English  ·  简体中文

Download  ·  First run  ·  Supported agents  ·  How it works  ·  Docs  ·  Research


dachshund memory library: memories from Claude Code, Codex, Cursor and other agents in one local feed

Is dachshund for you?

Every agent keeps its own partial model of you. A project rule lives in Claude
Code. A failed approach lives in Codex. A typography preference lives in Cursor.
Switching tools means losing context or repeating yourself.

dachshund takes the memory that already exists and puts it behind one quiet
surface:

  • Discover what is already on your Mac. Read known agent memory locations
    without rewriting the originals.
  • Keep one portable library. Store the unified result as plain Markdown, with
    source provenance retained.
  • See cross-agent echoes. When two agents remember the same thing, keep one
    memory and show every source.
  • Recall from any connected agent. A loopback-only MCP bridge exposes search,
    context, remember, forget and stats tools.
  • Stay out of the way. The whole utility lives in a compact macOS menu bar
    panel, with a full library window when you need it.

It is for you if you use more than one coding or desktop agent. If one agent
already remembers everything you need, you probably do not need dachshund.

Download

Platform Architecture Package Size Download
macOS Apple Silicon DMG About 120 MB Download v0.1.14

System requirement: macOS on Apple Silicon

Distribution note: The current MVP is ad-hoc signed and not notarized.
macOS Gatekeeper may require you to right-click the app in Finder and choose
Open on first launch.

All releases ·
SHA256SUMS

Install

  1. Download the DMG.
  2. Open the image and drag dachshund.app into Applications once.
  3. Launch dachshund from Applications.

The drag is only needed for the first install or recovery. From 0.1.3 onward,
a downloaded update replaces the installed app in place, relaunches it
automatically, and does not show another Applications drag target.

Optional: verify download integrity

The SHA-256 checksum confirms that the downloaded DMG matches the published
release. It is not an activation or license key and does not need to be entered
into the app.

cd ~/Downloads
curl -LO https://github.com/Ewen2015/dachshund/releases/download/v0.1.14/SHA256SUMS
shasum -a 256 -c SHA256SUMS

Build from source

Requirements: macOS on Apple Silicon, Node.js 22 or newer, and npm.

git clone https://github.com/Ewen2015/dachshund.git
cd dachshund
npm install
npm run build
npm start

To build the Apple Silicon DMG:

npm run dist:mac

First run

1 · Let it scan. On first launch, dachshund detects the agents installed on
this Mac and imports their existing memory into:

~/Documents/dachshund/memory/

2 · Connect an agent. Open Agents, turn on sharing for an installed
agent, then choose Connect MCP. Most agents receive only dachshund's named
MCP entry, with a backup before an existing config changes. App-managed
connectors, including Doubao, open a guided local setup instead.

Supported agents

Claude Code · Codex · Cursor · Gemini CLI · GitHub Copilot · OpenCode · Windsurf · Claude Desktop · WorkBuddy · Doubao · Trae · Hermes · Pi · ZCode · DeepSeek Harness

dachshund shows only the agents found on your machine. All fifteen integrations
currently have read-only importers. For Claude Desktop this covers local
Cowork Space memory; ordinary Chat memory stays on the account and remains
available to Claude over MCP. Doubao connects through a guided STDIO setup
because its custom connectors live inside the Doubao app rather than a stable
local JSON file. Doubao can then read and write dachshund's shared memory over
MCP; its account and per-assistant memories remain separate.

DeepSeek Harness is treated as event evidence rather than as a native memory
folder. dachshund derives explicit durable cues, imports compaction summaries as
History and ~/.dsh/AGENTS.md as rules, then exposes the shared library through
the managed Cordis MCP connection. Raw sessions and tool output stay evidence
and do not become memories.

See Agent integrations for exact paths and formats, and
Agent icons for brand asset provenance.

How it works

Claude Code  ─┐
Codex        ─┼─ read-only import ─→ ~/Documents/dachshund/memory/
Cursor       ─┤                         │
Other agents ─┘                         ├─→ SQLite FTS5 + BM25 index
                                        │
                                        └─→ 127.0.0.1 + bearer token
                                                   │
                                                   v
                                            MCP stdio bridge
                                                   │
                                        every connected agent

The Markdown files are the durable source of truth. SQLite is a rebuildable
index. The local HTTP service binds only to 127.0.0.1, and the MCP bridge reads
its short-lived token from an owner-only runtime file.

See Architecture for the process model, storage contract,
security boundary and failure modes.

What it does

Import first

Useful memory appears before you write anything. dachshund reads the formats
agents already produce, including Markdown, WorkBuddy notes, Trae session JSONL,
Hermes native memory entries, Pi summaries, ZCode SQLite rows, DeepSeek Harness
Zstandard logs, Codex memory files and readable Codex SQLite rows.

One memory, visible echoes

Identical memory from multiple agents becomes one item with every source shown.
You can trace where it came from without turning the library into a collection of
duplicate notes.

A real Mac utility

The tray icon, global shortcuts and 420 × 620 popover make the library ambient.
The panel opens on the display you are using and stays above the current
full-screen Space without switching screens. Use Cmd+Shift+M to minimize to
the menu bar or Cmd+Shift+D to toggle the compact panel. The full app can use
compact top tabs or an optional sidebar navigation inspired by Magpie's product
design; switch with the control beside the window buttons or Cmd+\. In sidebar
mode, Pulse sits under Insights. dachshund remembers the choice.

dachshund full app using the optional sidebar navigation, with the logo after the sidebar toggle

Pulse: the shape of your memory

Pulse is a memory-specific view, not a vanity usage dashboard. See which agents
are growing the library, how much knowledge is echoed across agents, how memory
is distributed across the five categories, and whether the library is getting stale. The same summary
is available from the menu bar panel.

dachshund Pulse tab with memory growth, cross-agent overlap and freshness

Updates in the same quiet surface

When a new GitHub Release is published, the menu bar panel and full app show an
update card with the version and release notes. dachshund downloads and verifies
the DMG, then exits while a small detached installer replaces the app in place.
The new version reopens automatically; the DMG is only opened as a fallback if
installation fails.

dachshund Agents page with real product marks and memory counts dachshund menu bar panel

MCP without a cloud memory layer

Connected agents can:

  • search shared memory;
  • request a prompt-ready context block;
  • write a durable memory;
  • archive a shared memory;
  • inspect library statistics.
Tool Purpose
memory_search Find relevant memories with source agents
memory_context Build a compact prompt-ready context block
memory_remember Save a durable shared memory
memory_forget Archive a shared copy without deleting the source
memory_stats Show memories, agents, sources and echoes

Plain files, real identities

The library is readable and portable without dachshund. Agent logos are cached
from the pinned MIT-licensed LobeHub icon set at build time; the app makes no
remote icon requests at runtime.

Privacy

  • No user account and no hosted service.
  • No analytics, telemetry or remote logging.
  • No API key, LLM or embedding call on the core path.
  • The only update network request is the public GitHub Releases endpoint.
  • Agent-owned files are read in place and never rewritten by import.
  • Archiving a shared memory does not delete the original agent file.
  • The HTTP service is loopback-only and requires a random bearer token.

dachshund is local by architecture, not by configuration.

Documentation

Product Product thesis, users, surfaces and boundaries
Memory taxonomy Native agent models, shared categories and classification rules
Design Visual system, layout, motion and accessibility
Pulse Memory visualization, metric definitions and design rationale
Architecture Processes, storage, MCP and security
Agent integrations Supported paths, formats and config writers
Agent icons Pinned icon sources and licensing
MVP Verification, checksums and known limitations
Development Data isolation, tests, screenshots and packaging
Updates Release contract, download verification and in-place installation
Research Local-first cross-agent memory landscape, with dated evidence and product implications
Memory taxonomy research Implemented classification model, native mappings, and clearly separated future proposals
Agent memory structures Per-agent storage models, persistent files, and dated evidence

Development

npm run typecheck
npm run test:agents
npm run build
npm run test:mcp
npm run demo:seed
npm run screenshot

Tests and screenshots use isolated DACHSHUND_HOME and DACHSHUND_LIBRARY
directories. They never modify a real user agent configuration.

Current limitations

  • Deduplication is exact, not semantic.
  • Library search is lexical, with no local embedding model.
  • Imports run at startup and on explicit scans; there is no live file watcher.
  • Source edits are not yet reconciled incrementally.
  • The build is Apple Silicon only and not notarized.
  • There is no cloud sync, team mode or multi-device history.

License

MIT

dachshund is an independent project. Product names and logos belong to their
respective owners. dachshund's menu bar, compact panel, and some interaction
patterns are inspired by Magpie; we are
grateful for that work. dachshund is not affiliated with Magpie and does not
reuse its brand assets.

Yorumlar (0)

Sonuc bulunamadi