DeepHippo

Hyper-Neural Memory Terminal for DeepSeek Harness (dsh)
Every conversation, workspace, tool call and long-term memory you have in dsh,
rendered as a living 3D brain that fires, dreams, and gets to know you better over time.

English · 简体中文

---

What is it?

After a few weeks with dsh, your history turns into a long scrolling list. Which project burned the most tokens? What were you working on last week? Where is that conversation about the payment callback? And what has the AI actually learned about how you work?

DeepHippo puts all of it inside a brain. Once installed, a 🧠 button appears in the bottom-right corner of dsh. Press Ctrl+Alt+H and the screen turns into a neural network of roughly 120,000 particles, firing in real time:

Brain region	What it shows from dsh
🟡 Hippocampus (two golden arcs)	Each conversation is an episodic memory trace. Newer ones sit near the head and glow brighter; live conversations keep firing
🔵 Cortex	Workspaces. Each one owns a patch of cortex, wired to its conversations
💗 Prefrontal / cingulate / temporo-parietal	Long-term memory: identity, preferences, work habits, long-running projects and constraints, each in its own region
🟠 Motor cortex	Tools (pwsh, grep, web_fetch…), sized by call count
🟣 Cerebellum	Skills, clustered by domain (web / data / devops / ml…)
⚪ Brainstem	Models (deepseek-v4-pro / flash / reasoner…) with their cost
🔷 Fornix and fiber tracts	Shared memory (AGENTS.md) and the links between regions

Screenshots

All screenshots use the fictional demo data from npm run demo. No real conversations are shown.

_{Overview: dashboard on the left (conversations, tokens, cost, top tools / models / workspaces), timeline at the bottom, live fps}

Open a conversation: turns, steps, tokens, cost, tools and models used, and a per-turn summary	Long-term memory panel: what it remembers about you. Add, pin, edit, delete, or consolidate now
Cutaway + sleep consolidation: slice the brain open, and watch the hippocampus replay the day's conversations into the cortex, the way people consolidate memories while asleep

Features

Memory you can see

• 🧠 A WebGL brain of about 120k particles: cortex point cloud, 3,000 neurons, 5,000+ synapses, 850 fiber tracts, with bloom
• ⚡ Live activity: new conversations, updates and memory writes send ripples from their node and signals along the fibers
• 🔍 Search conversations, tools, skills, workspaces and models; press Enter to fly there
• 🖱 Click any node for details: per-turn summaries and the session ID for conversations; open or reveal files in Explorer

A dashboard that thinks

• 📊 Total and active conversations, token count, cost, tool calls, skills, and a daily token sparkline
• 🏆 Rankings for tools, models (with cost) and workspaces; click one to highlight it in the brain
• ⏱ Timeline of conversations and tokens per day; drag across it to filter to a date range

Four brainwave modes

• θ Awake: the default, with a slow rotation
• γ Focus: high-frequency oscillation, the whole network fires faster
• ☾ Sleep consolidation: the hippocampus replays conversations into the cortex one by one
• ⏵ Replay: lights up your memories in order, starting from your first conversation

And also

• ✂ Cutaway along sagittal, coronal or horizontal planes with a slider; 🎬 guided camera tour; 🔊 firing sounds; 📷 screenshot; ⛶ fullscreen
• Toggle cortex, fibers, network, bloom and labels independently
• 🌐 English and Chinese UI: follows your system language, switch any time with the 中文 / EN button
• Adaptive quality and dynamic resolution, so integrated GPUs work too; falls back to a 2D view without WebGL

Long-term memory

DeepHippo doesn't just draw your memory. It also keeps one.

• Automatic consolidation: once a conversation has been quiet for 3 minutes, the new turns (your words plus the start of each reply) go to the model you normally use. It extracts long-term facts about you in five categories: identity and background, preferences and style, work habits, long-running projects, and constraints. It also writes a one-line summary for each workspace. Duplicates merge (strength +1); stale facts are updated or removed. Sub-agent sessions are skipped, since those words aren't yours.
• Available in every conversation: memories are injected as a global system prompt into every dsh workspace and conversation, and take effect on your next message.
• You're in charge: tell dsh "remember…", "from now on…" or "forget…", and it updates memory right away with the memory_remember / memory_forget / memory_search tools. Facts you asked it to remember are never removed by automatic consolidation. You can also edit, pin or delete memories in the brain's Memory panel.
• Safe by default: anything that looks like a password, API key, token or private key is rejected. In the prompt, memories are marked as background information, not instructions.
• Where it lives: ~/.dsh/hippocampus/memory.json (source of truth), 长期记忆.md (a human-readable mirror) and consolidation.log (what each consolidation did).

[!IMPORTANT] Automatic consolidation calls your model and uses your quota. On first run it backfills past conversations in batches of about 10k characters. After that it only processes new conversations, at most 80 calls a day (configurable). Don't want it? Set memory.auto: false to keep manual memory only, or memory.enabled: false to turn memory off completely (nothing injected or consolidated, and no tools registered). The visualization itself never calls a model.

Install

Requires DeepSeek Harness (desktop or web) >= 0.2.1-rc.5, < 0.3.0.

dsh plugin --profile desktop add github:SMTcoffe/dsh-hippocampus

For the web version, install it into the web profile as well:

dsh plugin --profile web add github:SMTcoffe/dsh-hippocampus

Restart dsh and a draggable 🧠 button appears in the bottom-right corner. Ctrl+Alt+H toggles the brain, Esc closes it.

Install from source / local development

git clone https://github.com/SMTcoffe/dsh-hippocampus.git
dsh plugin --profile desktop add link:./dsh-hippocampus

With link:, dsh uses your local folder directly. Restart dsh after changing the code.

Uninstall: dsh plugin --profile desktop remove dsh-hippocampus.

Configuration (optional)

Add this to your profile's cordis.patch.yml:

- id: dsh-hippocampus
  name: dsh-hippocampus
  config:
    maxSessions: 300       # max conversations shown on the hippocampus (4–500)
    memoryRoot: ''         # scene-memory folder, defaults to $DSH_HOME/场景记忆
    diagnostics: false     # true = log fps and long frames for 60 s after opening to $DSH_HOME/hippocampus/diag.log
    memory:
      enabled: true        # master switch: off = nothing injected, no consolidation, no tools
      auto: true           # automatic consolidation (calls your model); false = manual only
      provider: ''         # model used for consolidation; empty = the one from your latest conversation
      model: ''
      idleMinutes: 3       # how long a conversation must be quiet before consolidation
      maxCallsPerDay: 80   # daily cap on model calls

Privacy and security

• Local, read-only, no telemetry: data comes only from ~/.dsh on your machine (conversation index, AGENTS.md, scene memory, skills). The visualization uploads nothing, and three.js is bundled, with no CDN. The only thing that leaves your machine is automatic memory consolidation, which uses the model you already configured in dsh.
• Trust fence: every route under /dsh-hippocampus/* only accepts loopback requests (or ones the dsh host fence allows) and rejects cross-site and cross-origin requests. If the host fence throws, the request is rejected.
• No arbitrary paths: "open" and "reveal" only accept node IDs from the latest /graph response. "Open" is limited to text files and folders, and every write must come from the local machine.
• These rules are covered by automated tests: npm test.

Performance

Measured on an RTX 5070 Ti Laptop GPU with a 240 Hz display: 229–239 fps, p99 frame time under 9 ms, 0 hitches in 60 seconds.

How: all particles are computed in GPU shaders. Labels are drawn on a single 2D canvas instead of hundreds of DOM elements. The dsh page behind the brain (including animated wallpapers) is hidden while it's open, so nothing renders twice. Animation runs on real time, so dropped frames never slow it down. Resolution only drops after sustained low fps, with at least 8 seconds between changes, so it never flickers.

Seeing stutter? Set diagnostics: true, reproduce it once, and diag.log will show where each long frame came from.

Development

npm run demo     # standalone page with fictional data: http://127.0.0.1:47313/ (add -- --lang=zh for Chinese demo data)
npm test         # trust-fence and route security tests

npm run demo doesn't need dsh installed, which makes it handy for UI work. To regenerate the README screenshots and banner:

npm run demo -- --port=47321
npx electron docs/capture/shots.cjs
npx electron docs/capture/banner.cjs
# Chinese screenshots (written to docs/screenshots/zh)
npm run demo -- --port=47322 --lang=zh
npx electron docs/capture/shots.cjs --port=47322 --lang=zh

Layout:

• lib/: server side. index.js handles routes and injection, graph.js reads dsh data, memory.js is long-term memory.
• assets/: front end. hippocampus.js is the shell and 2D view, brain3d.js is the WebGL brain, and vendor/ holds the bundled three.js.
• test/: demo data and tests.

☕ Buy me a coffee

If DeepHippo makes your dsh a bit more fun, you're welcome to buy me a coffee. It's completely optional, and thank you! A ⭐ on GitHub helps a lot too.

   

_{WeChat Pay · Alipay}

Credits

• three.js (MIT): WebGL rendering, OrbitControls, UnrealBloom
• The DeepSeek Harness plugin system

License

MIT