{"schema_version":"newruntime-agent-readable-v0.2","type":"post","stable_id":"post:contextual-agent-memory-four-layer-system","slug":"contextual-agent-memory-four-layer-system","title":"A Vector Store Is Not An Agent Memory System","description":"Contextual AI separates working, procedural, semantic, and behavioral memory, with evaluation and provenance gates protecting every durable write.","retrieval_nugget":"Contextual AI separates working, procedural, semantic, and behavioral memory, with evaluation and provenance gates protecting every durable write. External memory is often more practical than training facts and behavior into model weights. It can be inspected, updated, validated, and carried across model upgrades.","status":"published","published_at":"2026-08-03","updated_at":"2026-08-03","record_date":"2026-08-03","date_kind":"published_at","topics":["memory","context-engineering","agents","evals"],"source_urls":["https://contextual.ai/blog/demystifying-agent-memory"],"visuals":[{"id":"contextual-agent-memory-four-layer-system","kind":"editorial-diagram","role":"hero","src":"https://newruntime.com/images/posts/contextual-agent-memory-four-layer-system.webp","alt":"Hand-drawn four-layer memory architecture separating current task context, tool-use playbooks, stable domain facts, and production behavior patterns, with provenance and evaluation gates controlling durable updates.","caption":"Different memories require different storage, retrieval, update, and validation rules; similarity search is only one component.","credit":"New Runtime synthesis from Contextual AI","source_url":"https://contextual.ai/blog/demystifying-agent-memory","generated_with":"gemini-3.1-flash-image","width":1600,"height":900,"legend":[{"label":"Working","description":"Current requests, evidence, tool results, and intermediate state share a finite context budget."},{"label":"Procedural","description":"Validated playbooks capture how this agent should search and use its tools."},{"label":"Semantic","description":"Stable domain facts and authoritative-source rules persist outside individual sessions."},{"label":"Behavioral","description":"Repeated production feedback changes prompts and response behavior after evaluation."}]}],"routes":{"html":"https://newruntime.com/posts/contextual-agent-memory-four-layer-system/","markdown":"https://newruntime.com/posts/contextual-agent-memory-four-layer-system.md","json":"https://newruntime.com/posts/contextual-agent-memory-four-layer-system.json"},"source_format":"markdown","next_reads":[{"type":"topic","path":"/topics/agents/","reason":"Explore the agents topic hub.","url":"https://newruntime.com/topics/agents/","title":"Agents - New Runtime","media_type":"text/html"},{"type":"topic","path":"/topics/context-engineering/","reason":"Explore the context engineering topic hub.","url":"https://newruntime.com/topics/context-engineering/","title":"Context engineering - New Runtime","media_type":"text/html"},{"type":"related_material","path":"/posts/patientagentbench-health-agent-safety/","reason":"Shares agents and evals.","url":"https://newruntime.com/posts/patientagentbench-health-agent-safety/","title":"PatientAgentBench Tests Health Agents As Workflows","media_type":"text/html"},{"type":"related_material","path":"/posts/ramp-agentic-risk-operations/","reason":"Shares agents and evals.","url":"https://newruntime.com/posts/ramp-agentic-risk-operations/","title":"Ramp Separates Agent Reasoning From Risk Decisions","media_type":"text/html"},{"type":"related_material","path":"/posts/firecrawl-mcp-context-efficient-web-tools/","reason":"Shares agents and context engineering.","url":"https://newruntime.com/posts/firecrawl-mcp-context-efficient-web-tools/","title":"Firecrawl MCP Turns Web Search into a Bounded Agent Capability","media_type":"text/html"}]}