Deep Dive into 5 Open-Source AI Tools: From Toolchains to System Ecosystems
> Originally published 2026-05-02. Translated and adapted.
> These five projects seem unrelated—an enhancement layer for Codex, a showcase for on-device phone AI, a memory system, a self-improving agent, and a screen-recording tool. But together they paint one picture: AI is turning from "conversational apps" into "engineering infrastructure."
1. Introduction: Five Slices, One Picture
| Project | Core Problem | Solution |
|------|---------|---------|
| OMX | Codex CLI lacks workflows and persistent state | Four-layer architecture: Skills + .omx/ state + tmux parallelism + AGENTS.md orchestration |
| Edge Gallery | Cloud AI depends on network; privacy and latency limits | LiteRT on-device inference + Gemma 4 + FunctionGemma function calling |
| MemPalace | Extractive memory loses information; API costs | Verbatim storage + zero LLM writes + progressive recall |
| Hermes Agent | Agents forget everything and can't learn from experience | Built-in learning loop + self-generated Skills + Honcho user modeling |
| OpenScreen | Post-production screen recording tools are expensive and closed | Electron + PixiJS + WebCodecs open-source pipeline |
Individually each solves a specific problem; together they point to a larger trend: the key to AI systems is no longer single-point capability (how strong the model is), but system-level orchestration, state management, and engineering.
2. Project Summaries
2.1 OMX: Turning Codex from a "Chatting Programmer" into an "Engineering Team"
In one sentence: OMX is a workflow enhancement layer for Codex CLI, not a replacement.
Core insight: Codex is smart but not engineered—no requirement-clarification flow, no persistent memory, no parallel work. OMX solves these with four keywords ($deep-interview → $ralplan → $ralph / $team) and persistent .omx/ state.
Most interesting engineering choice: tmux parallel workers + independent git worktrees. Each worker is a full Codex process in its own worktree; a leader coordinates via CLI API. This is the infrastructure that turns "conversational AI" into an "engineering pipeline."
Maturity: ~27k stars, iterated to v0.15.0 in two months, maintained by a Korean team, MIT license. Best for developers using Codex CLI on medium-to-large projects.
2.2 Edge Gallery: Google's "Showroom" for On-Device AI
In one sentence: Not a toy app—this is Google's on-device AI strategy showroom, packing LiteRT inference, the Gemma 4 model family, and the FunctionGemma function-calling specialist onto your phone.
Core insight: The value of on-device AI isn't "no network needed"—it's that in offline environments, privacy-sensitive scenarios, and real-time requirements, on-device is the *only* solution. LiteRT is a ground-up rework of TFLite unifying CPU/GPU/NPU abstractions; NPU inference can reach 25× CPU speed at 1/5 the power.
Most interesting features:
- FunctionGemma 270m: a 270M-parameter function-calling specialist that works offline on phones
- Mobile Actions: FunctionGemma-driven offline device control
- Thinking Mode: Gemma 4 supports "thinking" and "non-thinking" modes with explicit reasoning display
- Factual memory (
MEMORY.md) - Procedural memory (self-generated files in
skills/) - User profile (
USER.md, via Honcho) - 6 terminal backends (Docker/SSH/Daytona/Singularity/Modal/local); Modal's serverless sleep/wake means you can run a dormant agent free and pay only on wake
- 6 messaging platforms (Telegram/Discord/Slack/WhatsApp/Signal/Email)
- OpenClaw migration compatibility
- Recording: MediaRecorder API + WebAudio mixing
- Rendering: PixiJS (WebGL 2D engine) as a full video compositor
- Export: WebCodecs VideoEncoder + mediabunny MP4 muxer
- Timeline: dnd-timeline for drag editing
- OMX: layered orchestration +
.omx/persistent state - Hermes: learning loop + three-tier memory
- AiScientist: File-as-Bus + layered research teams
- Edge Gallery proves 7B+ parameter models run locally on phones
- MemPalace proves local storage can hit top retrieval performance at zero API cost
- Together they dissolve the assumption that "AI must depend on the cloud"
- OMX: The orchestration layer is their own; Codex CLI's reasoning is OpenAI's. tmux parallelism is clever engineering, not a new algorithm.
- Edge Gallery: LiteRT and Gemma are Google's; Gallery itself is "integration showcase" rather than fundamental innovation. But packaging scattered tech into a usable product is real engineering value.
- MemPalace: The verbatim philosophy is a real insight, but 96.6% on the benchmark is mostly thanks to all-MiniLM-L6-v2 + ChromaDB, not the palace metaphor. The palace is pretty folder structure.
- Hermes: The learning loop is compelling, but the "review every 15 tool calls" threshold is a heuristic, not optimal. Skill self-improvement needs stricter validation.
- OpenScreen: Web tech for video editing has costs—WebCodecs' software encoding speed is a hard limitation. It's a trade-off, not a silver bullet.
- OMX: give Codex an assistant team
- Edge Gallery: stuff models into phones
- MemPalace: store everything verbatim
- Hermes Agent: let agents keep work notes
- OpenScreen: do video editing with web tech
Maturity: 22.4k stars, officially maintained by Google, Apache-2.0, experimental Beta. For developers who care about on-device AI and privacy.
2.3 MemPalace: The Contrarian Bet on "Storing Everything Verbatim"
In one sentence: MemPalace bets that extractive memory methods may be wrong—store raw conversation verbatim and rely on retrieval.
Core insight: Industry consensus (Mem0, Zep, LangMem) uses LLM extraction and summarization, but extraction is lossy—you can't know what will be asked later. MemPalace achieves 96.6% LongMemEval R@5 with pure semantic search and zero LLM calls, far ahead of the then-leading extractive systems (Mem0 at ~49%) at zero API cost.
But don't be fooled by the "palace": an independent arxiv paper (2604.21284) critically concludes that the palace metaphor itself contributes ≈ 0 to retrieval performance (it's just ChromaDB metadata filtering); the real contribution comes from verbatim storage + zero LLM writes + minimal wake-up cost (~170 tokens). The paper calls it an "overhyped but genuinely architectural insight."
Maturity: ~48k stars (growing fast), MIT, Python 3.9+, ~300MB disk. Caveat: Mem0's new algorithm reached 93.4% in April 2026—verbatim's lead is shrinking.
2.4 Hermes Agent: An Intern Who Writes Its Own SOPs
In one sentence: Hermes's specialty isn't "a smarter model" but "distilling reusable skills from experience at runtime."
Core insight: Most agents are "interns with amnesia at the end of each day." Hermes's core mechanism: after every ~15 tool calls, the agent automatically reviews and generates a skill file (stored in ~/.hermes/skills/), invoked and improved automatically on similar future tasks.
Three-tier memory architecture:
Attractive engineering decisions:
Maturity: From Nous Research, MIT, Python 3.9+, supports 200+ models. For users wanting an agent that "gets better with use."
2.5 OpenScreen: Native-Grade Screen-Recording Post-Production with Web Tech
In one sentence: OpenScreen builds a complete post-production pipeline—recording, editing, zoom animations, motion blur, background replacement, annotations, speed control, multi-platform export—entirely with Electron + PixiJS + WebCodecs.
Core insight: Screen Studio ($29/month) proved "auto follow-mouse zoom" is a valuable category, but it's closed. OpenScreen's approach: implement 80% of core functionality with a web stack, free and open source.
Interesting tech stack:
Key limitation: WebCodecs encodes at only 1/7 of realtime (~5fps for 2K content)—1 minute of content takes 7 minutes to export. This is a hard boundary of the web stack: WebCodecs is a software encoder without hardware acceleration.
Maturity: MIT, very new project. For individual creators seeking a free Screen Studio alternative.
3. Horizontal Comparison
3.1 Technical Routes
| Dimension | OMX | Edge Gallery | MemPalace | Hermes | OpenScreen |
|------|-----|-------------|-----------|--------|-----------|
| Core engine | Codex CLI + Node.js | LiteRT + Kotlin | Python + ChromaDB | Python + LLM APIs | Electron + PixiJS |
| Persistent state | .omx/ filesystem | Local SQLite | Local files + SQLite KG | ~/.hermes/ filesystem | Project files |
| Parallelism | tmux + multi-worktree | Async concurrent execution | None (serial retrieval) | Sub-agents | None (serial rendering) |
| Cost model | Codex API only | Completely free (on-device) | Zero API cost (local embeddings) | LLM API costs | Completely free |
| Ecosystem | npm + Codex hooks | Google AI Edge | MCP server (29 tools) | agentskills.io + OpenClaw | Standalone app |
3.2 Maturity Matrix
| Project | Code quality | Docs | Community | Stability | Production readiness | |------|---------|------|------|--------|-----------| | OMX | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | Individuals/small teams | | Edge Gallery | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | Experimental Beta | | MemPalace | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | Production-ready | | Hermes Agent | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | Exploratory use | | OpenScreen | ⭐⭐⭐ | ⭐⭐ | ⭐⭐ | ⭐⭐ | Early prototype |
3.3 Target Users
| Project | Best fit | Worst fit | |------|--------|---------| | OMX | Engineers building projects with Codex CLI | Occasional Codex users writing small scripts | | Edge Gallery | Users who care about on-device AI, privacy, offline use | Heavy users needing cloud-scale model capability | | MemPalace | Privacy-sensitive, API-budget-limited users of Claude Code/Cursor | Scenarios needing complex relational reasoning | | Hermes Agent | Users wanting an agent that improves with use, multi-platform needs | Scenarios requiring strictly deterministic behavior | | OpenScreen | Individual creators wanting a free Screen Studio alternative | Teams needing fast export and professional video production |
4. Trend Insights: Three Ongoing Shifts
4.1 From "Model-Centric" to "Orchestration-Centric"
OMX, Hermes Agent, and AiScientist all point the same way: the bottleneck is no longer model intelligence but how multi-agent systems are orchestrated, how state is managed, and how long-horizon tasks stay on track.
Common theme: state continuity matters more than reasoning capability.
4.2 From "Cloud-First" to a Dual Track of "Edge + Local"
Edge Gallery (on-device inference) and MemPalace (local-first storage) represent another thread: not all AI needs the network.
4.3 From "Consumer Tools" to "Production Infrastructure"
OpenScreen's tech-stack choices (Electron + web tech for video editing) and OMX (turning conversation into an engineering pipeline) both show: "commoner technologies" like Web/Node.js/Python are eating into traditional professional software's territory. This isn't to say Electron is faster than native—it's that when the performance gap shrinks to acceptable range, development speed and ecosystem richness win.
5. The Feynman Question: What Are These Projects Really Doing?
Feynman would ask: "You all claim to solve big problems—but are you honest about what you built versus what you borrowed?"
Feynman's summary: Every project dances on the boundary between honest and dishonest. The most honest are those that proactively correct their own headline numbers (MemPalace v3.3.0 revised 100% to 98.4%); the least honest repackage standard techniques as revolutionary. Real value comes from people who acknowledge their boundaries.
6. Conclusion
None of these five projects claims to be "revolutionary"—they are all pragmatic engineering choices:
That's what's most worth watching in the 2026 open-source AI ecosystem: not the parameter race, but the engineering, orchestration, and infrastructuralization around models.
---
> Author's note: These five projects took about an hour to fully dissect with parallel sub-agents—three succeeded, two timed out, and I had to finish OMX and MemPalace myself. The Korean team behind OMX iterates fast: 12 minor versions in 2 months. I quite like MemPalace's verbatim philosophy, even if the marketing doesn't convince me—the palace metaphor is a gimmick, but the underlying idea of "throw nothing away" is right. Next time you tell me "this tool is nice," I'll first ask: is it honest?