Competitive Map
Category View
Garry's system spans three categories:
- Coding-agent workflow layer: gstack.
- Persistent agent memory layer: gbrain.
- Evaluation/proof layer: gbrain-evals.
The system competes less with one product than with the default way developers currently use AI: ad hoc prompts, editor autocomplete, scattered notes, and manual QA.
gstack Competitive Position
| Alternative | What They Provide | gstack Difference |
|---|---|---|
| Claude Code | Strong agentic coding environment | gstack adds reusable roles, gates, browser workflows, and sprint process |
| Cursor/Copilot | In-editor coding assistance | gstack is workflow/process-first, not autocomplete-first |
| OpenCode/Codex CLI | Agent execution in terminal | gstack provides an opinionated operating system for when and how to use the agent |
| Karpathy-style CLAUDE.md rules | General coding rules and preferences | gstack turns rules into slash-command workflows with outputs and gates |
| Conductor-style parallel sessions | Multi-session execution | gstack gives each parallel sprint a lifecycle, reviews, and stop conditions |
gstack Wedge
"AI coding is not safe because the model is smart. It is safe when the work process is encoded."
Strengths:
- Clear role metaphors.
- Fast first value.
- Works across hosts.
- Browser + QA + design capabilities.
- Strong founder narrative.
Risks:
- Skill sprawl can overwhelm new users.
- Claude-specific origins may create portability edge cases.
- Claims about productivity need durable methodology and caveats.
- Parallel sprint story depends on process discipline users may not yet have.
gbrain Competitive Position
| Alternative | What They Provide | gbrain Difference |
|---|---|---|
| Obsidian/Notion | Human-readable knowledge bases | gbrain is agent-operational with MCP, embeddings, source scoping, and skills |
| Mem0/Supermemory | API-first memory | gbrain is local/source-aware/code-aware and open-source |
| MemPalace/Hindsight/Mastra/Stella | Memory/retrieval systems and benchmarks | gbrain ties retrieval to an operational personal/work brain |
| Basic RAG over docs | Semantic search over files | gbrain adds pages, sources, graph/code edges, citations, maintenance, and workflows |
| Grep/read in codebases | Fast local search | gbrain adds symbol-aware retrieval and cross-session memory |
gbrain Wedge
"Your agent should remember the way an operator remembers: source-aware, cited, queryable, writable, and maintained."
Strengths:
- Strong schema and MCP surface.
- Source tenancy and trust policy.
- Code-aware retrieval.
- Skillpack turns memory into workflow.
- Public eval repo creates proof loop.
Risks:
- Operational complexity can exceed casual user tolerance.
- Setup paths are many: PGLite, Supabase, local MCP, HTTP OAuth.
- Memory quality depends on ingestion discipline and maintenance.
- Broad feature surface may dilute the core magical moment.
gbrain-evals Competitive Position
| Alternative | What They Provide | gbrain-evals Difference |
|---|---|---|
| LongMemEval | Public long-memory QA benchmark | gbrain-evals publishes runnable adapters and reports |
| LoCoMo/ConvoMem | Conversational memory tasks | gbrain-evals frames them as roadmap benchmarks |
| Product claims | Marketing assertions | gbrain-evals gives reproducible corpora, reports, and harnesses |
gbrain-evals Wedge
"Memory quality should be measured, not asserted."
Strengths:
- Public LongMemEval result.
- BrainBench fictional-life corpus.
- Adapter comparisons.
- Reports with caveats and reproducibility.
Risks:
- Benchmark quality must avoid appearing self-serving.
- Public comparison requires ongoing upkeep as competitors improve.
- Users may care more about lived workflow quality than benchmark numbers.
Strategic Takeaway
The system's real moat is not any individual skill, CLI, or table. It is the loop:
Workflow creates artifacts -> artifacts enter memory -> memory improves future workflow -> evals prove the memory works -> public proof attracts builders -> builders contribute back.
That loop is harder to copy than a prompt pack.
Deep Category Positioning
| Layer | Garry System | Competes Against | Wedge |
|---|---|---|---|
| Workflow/process | gstack | Claude Code alone, Cursor, Copilot, Codex CLI, OpenCode, Karpathy rules, Conductor | Encoded sprint process with roles, gates, QA, browser, security, and ship discipline. |
| Memory/runtime | gbrain | Obsidian, Notion, grep, basic RAG, Mem0, Supermemory, MemPalace, Hindsight, Mastra, Stella | Local/source-aware/agent-operational brain with MCP, skills, graph, citations, and code retrieval. |
| Proof/evals | gbrain-evals | LongMemEval, LoCoMo, ConvoMem, product marketing claims | Runnable public eval harness, benchmark reports, corpora, and caveats. |
gstack Versus Coding Agents
| Competitor | What They Own | gstack Counterposition |
|---|---|---|
| Claude Code | Strong coding agent runtime | gstack turns Claude into a role-based product/engineering org. |
| Cursor/Copilot | In-editor coding acceleration | gstack owns workflow before and after code: product framing, review, QA, ship, retro. |
| OpenAI Codex CLI | Independent terminal coding/review agent | gstack includes /codex as second opinion, making Codex a component inside the workflow. |
| OpenCode/Hermes/Factory/Kiro/Slate | Alternative agent hosts | gstack's host adapter layer says the host is replaceable; the method persists. |
| Karpathy-style rules | Lightweight agent discipline | gstack operationalizes rules as slash-command gates across a sprint. |
| Conductor | Parallel Claude sessions | gstack gives parallel sessions lifecycle, review routing, and stop conditions. |
Strategic wedge: the model is not the product. The encoded operating process is the product.
gbrain Versus Memory / RAG Systems
| Competitor | What They Own | gbrain Counterposition |
|---|---|---|
| Obsidian/Notion | Human-facing notes and knowledge bases | gbrain is agent-facing, queryable, writable, source-scoped, and MCP-native. |
| grep/ripgrep | Fast exact local search | gbrain adds hybrid retrieval, graph links, code edges, citations, and synthesis. |
| Basic vector RAG | Semantic retrieval over documents | gbrain adds source policy, pages, chunks, backlinks, typed links, timelines, skills, and maintenance. |
| Mem0/Supermemory | API-first memory products | gbrain is local/open/source-aware and built around operator-owned files. |
| MemPalace | High-performing memory benchmark posture | gbrain competes with reproducible LongMemEval numbers and a cheap deterministic retrieval headline. |
| Hindsight/Mastra/Stella | Memory/retrieval systems and benchmarks | gbrain frames itself as a full personal knowledge runtime, not only a benchmark adapter. |
Strategic wedge: agent memory should be durable, cited, source-aware, locally ownable, and maintained.
gbrain-evals Competitive Read
The LongMemEval report is unusually important because it turns memory from assertion into proof. The current story:
gbrain-hybrid: 97.60% R@5.gbrain-vector: 97.40% R@5.gbrain-keyword: 19.80% R@5.- MemPal raw: 96.6% R@5.
- MemPal hybrid+rerank held-out: 98.4% R@5.
- Supermemory and Mastra are caveated as QA accuracy, not retrieval recall.
The honest competitive story: gbrain is close to the best public memory benchmark numbers while keeping the headline retrieval path mostly deterministic and cheap. The exposed weakness is temporal reasoning, where the report says gbrain trails MemPal raw by 1.5 points and likely needs temporal extraction wired into ranking.
System-Level Moat
The real moat is the loop:
workflow artifacts -> memory ingestion -> better future workflow -> eval proof -> public credibility -> contributors/candidates/founders -> stronger workflow
That loop is harder to copy than any single skill file.
Strategic Risks
- Copy risk: gstack's skills can be cloned. The moat has to move toward integration quality, memory continuity, evals, and public operating proof.
- Complexity risk: gstack + gbrain + evals is a lot. The system needs tighter happy-path onboarding.
- Benchmark risk: memory vendors will optimize for LongMemEval. gbrain needs broader evals around real workflows, code retrieval, source attribution, and temporal memory.
- Host risk: Claude Code origin may leak assumptions into other hosts despite adapter support.
- Trust risk: memory systems touch sensitive data. Local-first and source-scoped OAuth help, but trust depends on visible defaults and auditability.
- Narrative risk: "810x" style claims are memorable but brittle. The safer durable claim is encoded workflow plus memory plus evals compounds operator leverage.
Takeaway For Ren/OpenClaw
Ren/OpenClaw should not copy gstack as a prompt pack. The strategic lesson is the integrated architecture:
- OpenClaw: orchestration and multi-agent runtime.
- Ren: opportunity intelligence, OSS contribution judgment, PR craft, maintainer reading.
- gstack-like layer: repeatable process skills and gates.
- gbrain-like layer: durable context, source-scoped memory, code-aware retrieval.
- eval layer: public proof that memory/retrieval/contribution workflows work.
The sharper Ren wedge: OpenClaw routes the work; Ren chooses the right work; memory preserves the learning; evals prove the system is improving.