hebel. today · 15 November 202619 / 79

Agentic Engineering · Context Engineering

Developers debate whether agent memory needs a full log or a curated desk

The case for append-only logs. Tun Shwe, who leads AI at Lenses, builds agent memory as an append-only event log to preserve the full history of evidence, hypotheses, and corrections. This structure allows engineers…

The case for append-only logs. Tun Shwe, who leads AI at Lenses, builds agent memory as an append-only event log to preserve the full history of evidence, hypotheses, and corrections (Packt Deep Engineering). This structure allows engineers at companies like Sugabots to inspect the reasoning behind an agent's conclusion when a reviewer challenges an assessment (Rak Siva from The New Stack). By keeping the original history, the team ensures that the next agent or engineer in a sequence receives the relevant evidence and corrections selected from that record.

The case for curated desks. Anthropic’s experiments with Pokémon Red showed that an agent recording only strategies, plans, and rules earned four badges, while the agent recording every encounter earned only one (Exploring AI from Exploring Artificial Intelligence). Researchers at Kanerva’s Desk argue that a fixed-size running state, which holds only key facts and open questions, outperforms a sprawling history that requires the model to sort through noise (The Brian Roemmele Report.). This approach mirrors findings from Distyl AI, where models performed better when fed a focused input of 300 tokens rather than a full conversation history of 113,000 tokens (AI Prompt Hackers).

The synthesis of approaches. The choice depends on whether the goal is auditing past decisions or maintaining current performance. While Sugabots preserves full records to allow engineers to contribute corrections to a shared investigation, the Kanerva model suggests that compacting the state is necessary to keep the agent from failing on its primary task.

Try today · 15 min

Initialize a Kanerva Desk for your current project.

Create a file named DESK.md with three sections: Key Facts, Open Questions, and Pointers.

You need

  • A text file or markdown document
  • A list of your current project's core assumptions

Steps

  1. Create a section for Key Facts and list at most five numbered, verified facts.
  2. Create a section for Open Questions and list what evidence would settle each one.
  3. Create a section for Pointers to your primary documentation or URLs.
  4. Paste this structure into your agent's system instructions or project context.
  5. Update one fact today by checking it against your primary source.

One email instead of your newsletter pile.

Forward the newsletters you already get (a Gmail filter does it) and tell Hebel your goals. Every morning it reads them with 20,000 others and sends one short email, with one small thing to do today. Fewer emails, not more. Free during beta.

Start my paper →
LinkedIn X

Sources

  1. Packt Deep Engineering on hebel. →Tun Shwe explains why stateless models require an external event log to maintain long-term memory for coding agents.

  2. Rak Siva, The New Stack on hebel. → ✎ Written by a personSugabots demonstrates how to separate working context from a shared investigation record for multi-agent teams.

  3. Exploring AI, Exploring Artificial Intelligence on hebel. → ✎ Written by a personAnthropic's Pokémon experiment illustrates why selective memory notes outperform exhaustive transcripts for agent performance.

  4. AI Prompt Hackers on hebel. →Distyl AI and Chroma report that models perform better on focused inputs than on massive, uncurated conversation histories.

  5. The Brian Roemmele Report. on hebel. →Kanerva's Desk provides a template for maintaining a fixed-size running state to improve retrieval accuracy.

More in Agentic Engineering

Verifying agent behavior through simulation versus runtime guardrails

Teams move bottleneck from code generation to verification

Coding agents trade re-planning for reusable workflows

Production agents require deterministic guardrails and context-aware triage

Context Engineering yesterday

Engineers pivot from prompt engineering to structured component manifests

Sign in to Hebel first

Likes, your collection and done Hebel need an account. We email you a six digit code; no password.