CRP vs RAG, MCP, LangChain & MemGPT¶

CRP is not a replacement for the tools you already use. It is a governance and context layer that sits beneath or alongside them, making every AI call safer, continuous, and auditable.

The stack¶

Layer	Technology	Role
Agent communication	A2A	How agents talk to each other
Tool access	MCP	How agents discover and call tools
Governance & context	CRP	How every AI call is grounded, governed, and continued
Inference	OpenAI, Anthropic, local models	Where generation happens

flowchart TB
    subgraph "Application / Agent"
        A[Your app or agent]
    end
    subgraph "CRP Governance & Context Layer"
        C[Context Envelope + CKF]
        S[Safety Policy + DPE]
        P[Provenance + Audit]
    end
    subgraph "Agent Infrastructure"
        M[MCP - tools]
        A2[A2A - agent chat]
    end
    subgraph "Inference"
        L[OpenAI / Anthropic / local LLMs]
    end
    A --> C
    C --> S
    S --> P
    A --> M
    A --> A2
    C --> L
    S --> L

High-level comparison¶

	CRP	RAG	MemGPT/Letta	LangChain/LlamaIndex	MCP	A2A
Context lifecycle	Full pipeline	Retrieval only	Virtual paging	Chain-based	None	None
Output continuation	Automatic	No	No	Manual	No	No
Structured extraction	6-stage	Chunk embedding	LLM-managed	Manual	No	No
Quality scoring	S/A/B/C/D	No	No	No	No	No
Provenance	Full DAG	Document source	No	No	No	No
Hallucination detection	13-stage DPE	No	No	No	No	No
Cross-session memory	CKF graph+vector	Vector DB	Archival storage	Manual	No	No
Safety enforcement	Protocol headers	No	No	No	No	No
Compliance evidence	Runtime export	No	No	No	No	No
In-window overhead	Zero	Low	High	Medium	Very high (schemas)	Varies

CRP and RAG¶

RAG retrieves relevant documents. It answers "what source material might help?"

CRP turns source material into structured knowledge and manages the entire call. It answers:

Which facts are most relevant right now?
Has the model already seen this fact?
Did the output actually use the facts?
Is the output hallucinated, contradictory, or fabricated?
How do we continue if the output truncates?
Where is the audit trail?

Use RAG for document retrieval. Use CRP to make every retrieval actionable, scored, and auditable.

Context management guide

CRP and MCP¶

MCP gives agents tools. It standardises how an agent discovers and calls external functions.

CRP governs the AI calls that use those tools. It adds:

Safety headers on every tool-planning and tool-result call
Context-source manifests so tool outputs are attributed
Safety budgets across multi-step agent loops
Audit trails for every decision

CRP also avoids the protocol-token explosion of sending full tool schemas in every window. Schemas are placed only in tool-selection windows, cutting agent-token overhead by ~90%.

Tools & Agents SDK

CRP and A2A¶

A2A lets agents exchange messages. It does not standardise risk state or context continuity.

CRP headers propagate across A2A hops. A downstream agent receives the same safety budget, chain ID, grounding score, and policy hash as the upstream agent, so governance is preserved end-to-end.

Multi-Agent Safety

CRP and LangChain / LlamaIndex¶

LangChain and LlamaIndex are orchestration frameworks. They make it easier to chain prompts, retrievers, and tools.

CRP does not replace them. It provides the protocol contract underneath:

A standard envelope format
Standard safety headers
Standard audit events
Standard compliance mappings

Teams can keep their LangChain/LlamaIndex application logic and route calls through the CRP Gateway.

When to use each¶

If you need...	Use
Document retrieval	RAG
Agent tool calling	MCP
Agent-to-agent messaging	A2A
Prompt chaining & orchestration	LangChain / LlamaIndex
Long-term virtual memory	MemGPT / Letta
Safety, governance, compliance, continuation, provenance	CRP

CRP is the governance layer¶

Think of CRP as the TLS of AI calls:

TLS encrypts and authenticates every HTTP request without changing the application.
CRP governs and contextualises every AI call without changing the model.

One base_url change gives you:

Unbounded context and continuation
Hallucination and injection detection
Tamper-evident audit trails
EU AI Act / ISO 42001 evidence

Why CRP?