ISSUE #38·106··📖 14 min
🔒Subscribe to listen to this issue

The Agentic Engineer

I read the repos so you don't have to.
Issue #38 | November 11, 2026
  • ANTHROPIC ON AGENT MEMORY: the definitive breakdown of how production agents should store and retrieve knowledge. Persistent memory improved Opus 5.5 mean reward from 0.655 to 0.838 on CL-Bench 1.0. claude-mem is at 115K stars.
  • BLACKSWAN: REFUTE BEFORE YOU ACT: drops in over any existing agent loop and adds +25.2 percentage points on AppWorld (no thinking mode). The Popperian framing is worth stealing even if you skip the full implementation.
  • Tool of the Week: Kiro bots. Named, long-running agents in Kiro Crew that own one standing job each. Triage issues, watch CI, prep release notes. Run on schedules, launch parallel worker sessions, and operate within per-bot tool permissions. Pro plan is $20/mo, no additional charge per bot.

Anthropic's Four-Layer Memory Architecture (And Why Dreaming Is the Part That Changes Everything)

Source: Anthropic Research | Published: November 6, 2026

Anthropic calls the process dreaming. Rather than relying exclusively on an agent's in-session memory updates, a separate background process can revisit transcripts, validate existing memory notes, correct mistakes, and build structured representations such as linked wikis. Those representations can then make useful knowledge available to future sessions without requiring the agent to reconstruct everything from conversation history.

Read the full issue

Subscribe to The Agentic Engineer to unlock this and every future issue.