title: ‘Tiered Cognitive Memory: Procedural Skills vs. Declarative RAG’ date: 2026-02-25 tags:
context_management options: token_size: 446 id: 26032 status: proposed superseded_by: null type: adr version: 1.0.0 birth: ‘2026-02-25’ authors:
name: Vadim Rudakov email: rudakow
.wadim@gmail .com description: 2026-02-25.
ADR-26032: Tiered Cognitive Memory: Procedural Skills vs. Declarative RAG¶
Date¶
2026-02-25
Status¶
proposed
Context¶
To achieve “Predictability from constraints”, we must solve the “Context Tax” problem where the agent is overloaded with reference data that obscures its core instructions. Additionally, this is the token usage problem ADR-26030 Stateless JIT Context Injection for Agentic Git Workflows.
Decision¶
Implement a Tiered Memory Architecture:
Skills: High-density procedural logic (Level 1/2 Disclosure).
RAG: Low-density declarative knowledge (JIT Retrieval).
Consequences¶
Need to be finished.
Alternatives¶
Pure RAG for Everything: Store logic/instructions in the vector store. Rejection Reason: Vector search is probabilistic and partial; the agent might retrieve the wrong “how-to” step, leading to catastrophic failure in structured tasks like YAML generation.
Pure Skills (No RAG): Try to fit all domain knowledge into the context window. Rejection Reason: The context window constrains lead to the agent cognitive degradation and too high token usage lead to high computational/financial costs. Historical logs and large specs must remain external.
References¶
Participants¶
Vadim Rudakov
Gemini (Principal Systems Architect)