November 6 agent guide: Claude Code 2.1.319 and agent memory

November 6, 2026 kept Claude in the agent control plane. Claude Code 2.1.319 makes long MCP tasks stop blocking the terminal turn and gives fleets a deny-instead-of-prompt path for unattended runs. The same day, Anthropic Engineering published “The components of agent memory,” a practical research brief on notes, salience, and dreaming. Bacancy Technology reads the pair as runtime hardening plus memory architecture, not a new Claude chat model.

What shipped on November 6

Claude Code 2.1.319: background MCP tasks and fail-forward denials

The headline for agent builders is MCP task backgrounding in the terminal. When an MCP server runs a tool call as a task, Claude Code moves it to the background right away and tells Claude the result when it finishes. That unblocks long browser, build, or research MCP calls that previously parked the agent. CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND restores synchronous waiting. CLAUDE_CODE_DISABLE_MCP_TASKS=1 disables MCP tasks entirely.

permissions.denyInsteadOfPrompting

The new setting denies a tool call that would wait on a permission prompt and continues the turn in terminal sessions, subagents, and teammates. It mirrors --permission-prompts none for claude -p. Allow rules, permission mode, auto mode’s classifier, and PermissionRequest hooks still decide first. A streak of consecutive denials ends the turn. Bacancy Technology recommends this for CI and SDK hosts that must never hang on a human prompt, after allowlists are proven.

Gateway and fleet telemetry

• /usage can show spend for today, this week, and this month for developers behind a Claude apps gateway with no spend limit when the gateway is current.
• admin.pooled_conflict_rule and admin.pooled_priority_order control which pooled group budget pays first when a developer is in several groups.
• Optional login_notice shows an operator message on gateway verification and Claude Code sign-in.
• CLAUDE_GATEWAY_LOG_FORMAT=json emits structured operational logs for collectors.
• /tasks gains an “Earlier in this session” row for finished subagents that can be reopened.

This release continues the November 3 theme: control plane over model flash. Keep October 30 write blocks, October 31 layered policy, and 2.1.317 MCP link cards in place while rolling 2.1.319.

Anthropic on the components of agent memory

Anthropic’s November 6 engineering post defines agent memory as knowledge Claude gains from experience and retains between sessions as readable context. It excludes weight updates, single-conversation in-context learning, and compaction summaries that replace older turns. The practical stack is memory notes written during a session, salience judgment about what belongs in those notes, and dreaming: offline reorganization of logs into cleaner notes and representations.

What the evaluations show

• On Continual Learning Bench database tasks with a persistent memory directory, newer Claude models close the loop: validated rules, checked facts, and warnings about traps, instead of scattered untested guesses.
• On Pokémon Red, Claude Sonnet 3.7 wrote hundreds of files and stalled on stale state, while Claude Fable 5 kept a small strategic folder (global.md plus plans) and progressed further.
• Dreaming is framed as sleep-time compute that merges duplicates, replaces contradicted entries, and shapes what future sessions preload.

For production agents, Bacancy Technology’s takeaway is architectural: store durable rules and verified facts, not raw screen state the agent can re-observe; run consolidation offline; and measure memory by task progress, not file count.

Builder checklist

1. 01
   Upgrade to 2.1.319 and retest one long MCP tool

   Confirm background completion messages arrive, then decide whether any critical path needs CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND.
2. 02
   Enable denyInsteadOfPrompting on headless hosts

   Prove allowlists first, then turn the setting on for CI and SDK so runs fail forward instead of hanging.
3. 03
   Rewrite the agent memory folder contract

   Adopt a small global.md-style dashboard, ban redundant state dumps, and schedule a dreaming or consolidation job between sessions.

Conclusion

November 6 is Claude Code hardening plus memory design. Background MCP tasks and denyInsteadOfPrompting keep fleets moving. Anthropic’s memory components post gives teams a shared vocabulary for notes, salience, and dreaming. Ship the CLI upgrade, then fix the memory layer most agents still skip.

Related reading

• November 5 GPT-6 Cyber and Live Apps: /blog/november-05-gpt6-cyber-cursor-live-apps-2026
• November 3 Claude Code 2.1.317: /blog/november-03-claude-code-2-1-317-mcp-security-2026
• November 2 message threads: /blog/november-02-claude-message-threads-glm-legal-2026
• Three paradigms of LLM memory: /blog/three-paradigms-of-llm-memory-implicit-explicit-agentic
• MCP topic map: /topics/mcp

Need Claude fleets and agent memory reviewed together?

Bacancy Technology reviews Claude Code permissions, MCP task policy, gateway budgets, and agent memory architecture for production teams. Consulting: /consulting/ . Contact: /contact/ .

Sources: Claude blog ; Anthropic .