AMP (Agent Memory Protocol) is a six-verb specification for agent memory: encode, recall, forget, consolidate, pin, stats. Brain OS is a working MCP server with typed operational state. They are different kinds of thing, one is a protocol, one is a tool with an opinion.
AMP is an early-stage open specification that tries to standardize memory primitives across backends. Its six verbs (encode, recall, forget, consolidate, pin, stats) are oriented to the fuzzier, semantic side of memory: encode something, retrieve it later, consolidate over time. An AMP-compatible backend could plug into many agent frameworks.
AMP is a protocol, not a product. It does not have an opinion about what to remember, that's up to the implementation. Its strength is interoperability.
Brain OS is the opposite trade-off. It is a working tool with strong opinions about what to remember: decisions with reasons and rejected alternatives, plans with phases, blockers, focus, momentum, recurring patterns. The MCP tools (focus_get, decision_log, plan_read, pattern_detect, semantic_recall) encode those opinions. It is not a generic memory interface; it is a structured operational-state interface.
The difference: AMP asks "how should memory primitives be standardized?" Brain OS asks "what operational state does an agent actually need to do work?"
npx.brain/ shared by Claude Code, Cursor, Zed, GitHub Copilot, any MCP clientdecision_check returns conflict/caution/clear with prior reasoningyou're designing a memory abstraction across multiple agent backends, or evaluating where the ecosystem might standardize over time.
you want to install something today that gives your agent operational state (decisions, plans, blockers, focus) without designing a protocol or wiring a backend.
These are different kinds of artifact entirely. A protocol describes the shape of an interface; Brain OS is a concrete implementation with a specific point of view.