status: active
timestamp: 2026-07-02
tags: [agent, memory, skills, portability, strategy, karpathy]
Own the memory, rent the intelligence
Memory + skills + orchestration are portable and belong to you. Intelligence (frontier models) is rented — swap providers freely. Never couple your knowledge to one vendor's tool.
Own the memory, rent the intelligence
The principle
Memory belongs to you. Intelligence is rented.
- Memory =
knowledge/OKF bundle +~/.claude/projects/*/memory/per-session facts. - Skills =
repos/own/infra/agent-skills/junctioned into every fleet agent’s skills dir. - Orchestration =
.mcp.json+AGENTS.md+ reusable workflows inchirag127/workflows. - Intelligence = whichever frontier model is best/cheapest/available this week — Opus 4.8, Fable 5, Sonnet 4.6, Haiku 4.5, GLM 5.2, Gemini, or a local model.
Models get locked (Fable pulled last week), vetted-only (GPT 5.6), deprecated (any older tier), or price-gouged mid-quarter. Your memory + skills + orchestration survive every provider swap.
What this means in practice
| Layer | Rule |
|---|---|
| Memory | Every durable fact goes to knowledge/ OKF, never in-chat-only or vendor-owned “memory feature” |
| Skills | Every reusable operation goes to repos/own/infra/agent-skills/<skill>/SKILL.md |
| MCP config | Single source of truth at .mcp.json, synced to all agents via scripts/sync-mcp-configs.mjs |
| Agent rules | Single source of truth at AGENTS.md, auto-imported into every agent per agent-fleet-parity |
| Model choice | /model per-task. Never build features that require a specific model. |
Anti-patterns
- ❌ Storing a decision in ChatGPT’s persistent memory instead of
knowledge/ - ❌ Building a workflow only Claude Code can run (uses tools no other agent has)
- ❌ Committing memory-file paths that assume a specific harness location
- ❌ Setting
ANTHROPIC_MODEL=claude-opus-4-8in a place where switching is painful - ❌ Losing chat history to “the search doesn’t find it” — that’s exactly the failure mode; keep durable facts in
knowledge/, not in chat
The two anchors
-
Frontier providers get locked. Fable disappeared 2026-06-27, came back with 6-day promo period. GPT 5.6 shipped to vetted shops only. GLM 5.2 launched open-source in the same week. When any single-provider dependency exists, migration cost is high. When memory is portable, migration is a
/modelcommand. -
Agents will one-shot most workflows. Cost of building the “own memory” layer dropped ~5× in 6 months. Claude Code + Codex + open-source agents can now write the SQL + configs + skills for you. The technical build barrier is nearly gone; only the human parts (accounts, permissions, approval, taste) belong to you.
The “prove it” corollary
An agent that acts without evidence is out of policy. Every agent operation must:
- Cite what it read (files, URLs, prior context)
- Point to what it changed (paths, commit SHAs, PR URLs)
- Say plainly when something is unverified
Ties to karpathy-guidelines §Goal-Driven Execution and the “make it prove it” habit in fable-5-prompting.
Cross-refs
agent-fleet-parity— same rules + skills across all 6 agentsmcp-config-single-source-of-truth— MCP portabilityglobals-derived-from-workspace— workspace canonical, globals are cachekarpathy-guidelines— assumptions before code, verifiable goalsfable-5-prompting— “make it prove it” verification habit- Source: Nate B Jones YouTube 2026-06 + Nikita’s Lemonade insurance story