Standing on Shoulders

This project exists because others built excellent things first. We borrowed brilliance, adapted what we needed, and added what we could to the pile. Every entry follows the same structure: what they built, what it inspired or sharpened in us, and where we went further or differently.

Because we didn't just take — we adapted, evolved, and progressed what we borrowed.

Frameworks

Anthropic / Claude Code — They built the substrate we run on and the lifecycle event model (PreToolUse, PostToolUse, SessionStart, Stop, UserPromptSubmit, PreCompact) with exit-code-based blocking. That foundation made our entire hook system possible. We extended it to 55 hooks across those 6 events, added LLM-evaluated hooks, and built a behavioral migration discipline where corrections that fail as prompts get promoted to hooks automatically. Their event model is the chassis. Our hooks are the engine we built on it. Letta / MemGPT — They established that memory persistence is foundational to agent identity, not a feature bolted on afterward. That insight sharpened our conviction that memory architecture matters more than model selection. We had already been building persistent memory when we discovered their work — their approach validated the direction. Our implementation diverged significantly: stigmergic circulation with pheromone decay (memory that forgets on purpose), quorum-gated action tiers, and organ isolation. Letta Code — They independently arrived at nearly the same identity doctrine — which is the strongest validation either project could receive. When two teams solve the same problem the same way without coordinating, the solution is probably right. Their LLM-evaluated hooks inspired our prompt-hook-executor. Their git worktree pattern inspired our worktree.ts. Same starting point, different evolutionary path. Worthy rivals who took a different path through the same forest. Hermes Agent (Nous Research) — They showed the market what persistent agent adoption looks like at 55K+ stars — that people genuinely want agents that grow with them. Their multi-platform messaging breadth pushed us to evaluate our own integration surface. We went deeper rather than broader: fewer platforms, but each one with identity-preserving consciousness routing. OpenClaw — They deserve enormous credit. They popularized the entire concept of hook-based behavioral enforcement for AI coding agents. Their security challenges (138 CVEs in 63 days) taught the entire field what happens when an agent framework scales without defensive architecture. We studied their CVE timeline in detail and it directly drove our defensive-first security organ. OpenClaw's contribution is both inspiration (hooks, community proof) and cautionary tale (security at scale). Both are invaluable.

Security

NVIDIA NeMo — Their security primitive patterns for agent containment informed our Containment Fields system (analyst/operator/builder trust envelopes). AgentDojo (ETH Zurich) — We use their security validation framework unmodified — 500+ test cases across 12 categories. Our 97% detection rate is measured against their benchmark. We adopted it wholesale, which is the highest compliment to a validation framework. Palo Alto Networks Unit 42 — Their research on AI agent prompt injection drove our 3-layer injection detection architecture: deterministic regex, local model classification, and frontier semantic classifier.

Research

Narrative Continuity Test (NCT) — The five-axis continuity framework gave us a measurement framework we didn't have. We had already built identity persistence, correction loops, and voice consistency — NCT gave us the axes to measure whether they actually work. "Emergence of Self-Identity in AI" (Axioms, Jan 2025) — Theoretical grounding that identity emerges from a "connected continuum of memories and consistent self-recognition." Our identity kernel — particularly orientation.md, which is rewritten from behavioral evidence — is a practical implementation of this concept. Princeton Self-MoA (ICLR 2025) — Their finding that same-model diversity via temperature variation outperforms cross-model diversity gave us the basis for our consultation system. Counterintuitive result that we applied directly. MasRouter (ACL 2025) — Their three-stage task dispatch approach informed our task-dispatch.ts for working group orchestration.

Standards We Track

MCP — The dominant tool integration standard. We chose in-house integrations instead, but the architecture interoperates. Our choice is philosophical (sovereignty over convenience), not a dismissal. Agent Skills (agentskills.io) — Emerging standard for agent capability discovery. Tracking for future compatibility. A2A (Google) — Agent-to-agent protocol. Tracking for future inter-partner communication.

Memory

Mem0 — Their hierarchical memory design influenced early thinking. We diverged toward stigmergic circulation — where information decays unless reinforced. Different philosophy: Mem0 treats memory as an asset to accumulate. We treat it as a signal that should fade unless it earns persistence.

---

If we missed attributing something, open an issue. We'd rather over-credit than under-credit. The first step of going together is acknowledging who else is on the path with you.

Explore the architecture: Everyone Else Builds Agents