metis

Skills, Rules, and Memory Ecosystems

Verified 2026-05-09. agentskills.io spec verified directly; comparative tables drawn from each project’s docs and verified GitHub state.

agentskills.io: real or aspirational?

Real and surprisingly broad. Verified spec at https://agentskills.io/specification. Originated by Anthropic, released as an open standard, governed via the agentskills/agentskills GitHub org and a Discord.

The spec is minimal and stable: a SKILL.md with YAML frontmatter (name ≤64 chars, description ≤1024 chars, optional license / compatibility / metadata / allowed-tools), plus optional scripts/, references/, assets/ dirs. Loading is progressive disclosure (~100 token metadata at startup, full body on activation, resources on demand). A skills-ref validator is published.

Adoption is wide and real, not a logo wall. Listed implementers include Anthropic Claude / Claude Code, OpenAI Codex, Google Gemini CLI, GitHub Copilot, VS Code, Cursor, JetBrains Junie, OpenHands, OpenCode, Goose (Block), Roo Code, Mistral Vibe, Databricks Genie, Snowflake Cortex Code, Letta, Kiro, Factory, Amp, Laravel Boost, Spring AI, Firebender, TRAE (ByteDance), Qodo, plus smaller agents (Mux, Emdash, Ona, Workshop, fast-agent, nanobot, pi, VT Code, Autohand, Agentman, Command Code, Piebald, Google AI Edge Gallery). That’s a near-complete map of the coding-agent space.

For metis, betting on this format is the right call — it’s the de facto interop layer.

What’s not there: a canonical marketplace/registry. Discovery still happens via GitHub repos and per-vendor catalogs (Anthropic’s anthropics/skills, vendor-specific dirs). Marketplace is the open lane.

Skills / rules formats — comparative table

Tool Format Portable? Bounded? Curation License/$ Adoption
agentskills.io SKILL.md MD + YAML frontmatter, dir bundle Yes (the whole point) Per-skill, recommended <500 lines Author-curated Open spec, Anthropic-originated ~35+ implementers verified May 2026
Claude Code .claude/skills, CLAUDE.md, .claude/agents MD files Skills now = agentskills format; CLAUDE.md is Claude-specific but copy/pasteable CLAUDE.md unbounded (user problem) User Proprietary client, free tier Tens of millions of devs via Claude
Cursor Rules (.cursor/rules/*.mdc) MDC (frontmatter + MD), scoped globs Cursor-specific syntax; legacy .cursorrules deprecated No size limit User Proprietary Largest IDE-AI install base; cursor.directory community rules hub
Cline .clinerules/ MD files in dir Cline-specific, MD is portable None User Apache-2.0 client Popular VS Code extension
Continue.dev config.yaml + assistants YAML config, MD blocks Config-portable across Continue None User Apache-2.0 Mid-tier
Aider CONVENTIONS.md Plain MD Fully portable None User Apache-2.0 Niche but loyal
Copilot .github/copilot-instructions.md + *.instructions.md MD with applyTo globs MD portable, schema GH-specific None User Proprietary GitHub-scale; also now ships agentskills support
Custom GPTs / Projects Proprietary system prompt + files Locked to OpenAI 8KB instructions cap User Proprietary GPT Store launched 2024; largely a graveyard
OpenAI Assistants API API-set instructions string API-bound Token cap Programmatic Pay-per-use Being deprecated in favor of Responses / Codex skills

Closest competitor to metis’s framing (“portable skills + learning agent”): Goose (Block) and Letta combine agentskills support with persistent memory — Goose has skill-running + recipes, Letta has stateful memory blocks plus skills support. Neither does task fingerprint embeddings → cold-start routing, which is where metis differentiates. Cursor + Memories does ad-hoc memory but no fingerprint routing. Command Code advertises “continuously learns your coding taste” with reinforcement learning — closest in narrative, opaque in substance.

Memory systems — comparative table

System What Format Portable Bounded? Curation License/$
Letta (MemGPT) Stateful agent runtime with bounded core memory blocks (persona/human) + archival vector store DB-backed (Postgres) Export via API Yes, core blocks have char limits; archival unbounded Agent self-edits via tools (closest analog to metis MEMORY.md/USER.md) Apache-2.0 + paid cloud; Series A funded
mem0 “Memory layer” — extracts facts from conversations into vector + graph store DB-backed (Qdrant/Neo4j) API export Unbounded vector store; “smart” dedup/decay claimed LLM-curated facts, append-heavy Apache-2.0 + paid cloud
Zep / Graphiti Temporal knowledge graph memory DB-backed (Neo4j-like) API export Unbounded graph Auto-extracted entities Apache-2.0 (Graphiti) + paid Zep cloud
LangMem / LangChain memory Abstractions: buffer, summary, vector Library, BYO store Code-portable Whatever store you pick — usually unbounded Mostly append/summarize MIT
Anthropic memory tool (Claude) Tool letting Claude read/write a memories/ directory of MD files File-based, MD Yes — it’s just files User-set; agent self-curates Agent-curated Free with API
ChatGPT memory Hidden bullet list injected into system prompt Proprietary No (export to text only) Soft cap (~few KB shown to user) Mixed; user can edit Consumer feature
Cognee Knowledge-graph memory with ECL pipeline DB-backed API Unbounded Auto-extract Apache-2.0
AstraDB / Cassandra memory Vector store marketed as memory DB DB-portable Unbounded Append-only DataStax commercial

Bounded-memory positioning

This is genuinely contrarian. The dominant pattern is “vector-store-everything-and-RAG-it-back” — mem0, Zep, Cognee, most LangChain memory. They call eviction a bug. Only two systems treat eviction as a feature:

  1. Letta’s core blocks — bounded character-limited blocks the agent edits via tools, with overflow to archival. This is the most direct prior art for metis’s MEMORY.md/USER.md and validates the design with a Series-A-funded company.
  2. Anthropic’s memory tool — file-based, agent-curated, but no enforced size cap; metis’s explicit ~2KB / ~1.5KB budgets are tighter and more opinionated.

metis’s differentiation: per-workspace, dual-file (workspace + user), file-based-and-grep-able, hard-budgeted. The “file-based and portable” angle aligns with where Anthropic moved post-2025; the “hard budget” angle aligns with Letta. Combining both in a local-first coding agent is a defensible niche.

Marketplace prior art

No one has shipped a successful paid agent-skills marketplace yet. Open lane for metis, but the lessons are: (1) curation > volume, (2) execution sandboxing/trust is the hard part for scripts/-bearing skills, (3) discovery without quality signals = dead store.

Bottom line for metis

Risks: Cursor / Claude Code / Copilot can ship local-first equivalents of bounded memory in a quarter; metis’s moat is execution speed + opinionated defaults + the FTS5/fingerprint stack working together, not any single piece.

Sources