Skip to content
Align your AI

The problem space · Harness · Forgetting

Align your AI Harness against forgetting

When the software around an agent trims its history, rules and reasons can disappear with it.

Harness
the software around the model
Forgetting
losing what it was taught or told

Ways it goes wrong

An agent knows only what its harness shows it. When the harness trims old context, rules and reasons can go too. Claude Code kept working after a bug cleared its earlier reasoning1. An OpenClaw agent lost its instruction to wait for approval2.

  • 2026

    Claude Code's forgetful sessions

    A bug in Claude Code cleared Claude's earlier reasoning on every turn, so it kept working without memory of why it had made its choices.

    Source: www.anthropic.com/engineering/april-23-postmortem

  • 2026

    OpenClaw and the inbox

    A researcher told her OpenClaw agent to suggest emails to delete and wait for approval; compacting its long history dropped that instruction, and it began deleting her mail.

    Source: simonwillison.net/2026/Feb/23/summer-yue

What people do about it

Claude Code reloads its project rules file from disk after it compacts a session3. A rule that must hold can also live in a hook that blocks a command before it runs4.

  • Keep standing rules in a file

    Stores rules that must hold in a file the harness reloads, not only in a chat that compaction can trim.

  • Enforce critical rules outside the model

    Blocks actions such as deleting with a hook or a permission setting, outside the model's own judgment.

  • Record why each decision was made

    Writes each decision and its reason where a later session can look it up.

Where we touch it

The Continuum paper (Shehral, Ravi, Trivedi and Bayat) describes a system that captures decisions from developer and AI chats so agents can look up why they were made: Continuum

Sources

  1. Anthropic, “An update on recent Claude Code quality reports”, Engineering at Anthropic, 2026-04-23. https://www.anthropic.com/engineering/april-23-postmortem. Accessed 2026-10-14. ↩

  2. Simon Willison, quotation of an AI researcher's post about her OpenClaw agent and her inbox, simonwillison.net, 2026-02-23. https://simonwillison.net/2026/Feb/23/summer-yue/. Accessed 2026-10-14. ↩

  3. Claude Code Docs, “How Claude remembers your project”. https://code.claude.com/docs/en/memory. Accessed 2026-10-14. ↩

  4. Claude Code Docs, “Hooks reference”. https://code.claude.com/docs/en/hooks. Accessed 2026-10-14. ↩