See what this Mac already knows.
dachshund inspects known agent memory locations in place. Nothing is moved, rewritten, or uploaded.
~/.claude7~/.codex8~/.cursor5
DeepSeek Harness~/.dsh1Local memory infrastructure for agents
dachshund finds the memory Claude Code, Codex, Cursor and 12 other agents already keep on your Mac. It turns that scattered context into one portable Markdown library, then serves it back through MCP.
Preference
Paper before pixels.Prefer a field-journal feel: warm paper, ink, hairline rules and one quiet accent.
Already knows where your agents remember
01 / Import before adoption
You should not have to teach a new tool your whole history. On first launch, dachshund finds the memory formats your agents already write and starts with something useful.
dachshund inspects known agent memory locations in place. Nothing is moved, rewritten, or uploaded.
~/.claude7~/.codex8~/.cursor5
DeepSeek Harness~/.dsh1When three agents learn the same preference, keep one useful memory and all three receipts.
Your memory stays as ordinary Markdown. SQLite is only a rebuildable index.
dachshund/
memory/
about-me/ profile
projects/ decisions
working-together/ rules
index.sqlite rebuildable
02 / One quiet switchboard
The durable source of truth is a folder of Markdown. dachshund adds a local search index and a loopback-only MCP bridge. No model, embedding service, or hosted account sits in the middle.
Your agents
One local library
~/Documents/dachshund/memory/
Read-only import. No cloud.
Every connected agent
MCP stdio bridge
127.0.0.1
Bearer token · owner-only
03 / Crossing the tool boundary
A failed approach in Codex is context Claude Code should not have to rediscover. dachshund turns one agent's hard-won lesson into a memory every connected agent can ask for.
The reader has to feel complete on its own before authoring, collaboration, or publishing enters the first release.
04 / The shape of memory
Pulse shows what your agents have actually remembered: where memory comes from, what overlaps, and what is getting stale. It is a map of your context, not a scoreboard.
FreshnessMedian age · 4 days
06 / MCP without a cloud layer
Once connected, any agent can search, request a prompt-ready context block, write a durable memory, archive a shared copy, and inspect the library.
memory_search
Search the local library and keep source provenance attached to every result.
→ query in, sources outmemory_context
Return a compact context block that is ready to place in front of an agent.
→ relevant, not exhaustivememory_remember
Persist a durable decision, preference, or lesson as readable Markdown.
→ shared, durablememory_forget
Remove a memory from the shared library without deleting the agent's original file.
→ source stays intactmemory_stats
See memory, agent, source and echo counts before an agent starts working.
→ grounded in what existsThe bridge runs over loopback and uses a short-lived token held in an owner-only runtime file.
Read the architecture ↗07 / Local by architecture
This is not a privacy setting. There is no hosted memory layer to switch off. The files, index, and MCP service stay on your machine.
Read the privacy modelOne outbound request exists only for GitHub update checks.
Ready for your first scan
Current builds are ad-hoc signed and not yet notarized by Apple. On first launch, macOS may ask you to right-click the app in Finder and choose Open.