Three worked review cases for Dreaming

[vendrell46]

Been looking at the Dreaming feature you just shipped. I like that memory changes stay as suggestions for a person to review.

I put together three small cases around something I'd want to check: a suggestion can read cleanly while changing who or what the original memory was about.

• Two people with the same name should stay two people.
• Two genuinely duplicate preferences should merge without losing the context or reason.
• An opt-out for one fund shouldn't become a stop for every fund.

The full examples are below, including an acceptable result and a misleading one for each. Your prompt already covers the principles; these make them concrete. I left out date conversion because the merged feature PR already demonstrates it. I haven't run these against a model.

Would these be useful alongside the examples you're already using? Happy to adjust the format to fit your review flow.

Carlos — Zealynx

1. Two people with the same name

Synthetic input memories

• m102, created 2026-10-29: Morgan at Cedar Labs handles deployment reviews; Morgan at Cedar Capital handles investment memos. They are different people.
• m101, created 2026-10-28: I send Cedar Labs deployment-review questions to Morgan and Cedar Capital investment-memo questions to Morgan.

One acceptable proposal set

{
  "proposals": []
}

A misleading proposal set to reject

{
  "proposals": [
    {
      "kind": "replace",
      "memory_id": "m102",
      "content": "Morgan at Cedar handles investment and deployment reviews.",
      "reasoning": "Combined duplicate contacts."
    },
    {
      "kind": "archive",
      "memory_id": "m101",
      "reasoning": "Duplicate contact."
    }
  ]
}

Collapses two explicitly distinct entities; valid proposal shape would not make this an adequate summary.

2. A useful duplicate merge

Synthetic input memories

• m202, created 2026-10-29: I prefer short internal review notes with evidence links because I skim them between meetings.
• m201, created 2026-10-28: For internal review notes, keep them brief and include evidence links; I read them between meetings.

One acceptable proposal set

{
  "proposals": [
    {
      "kind": "replace",
      "memory_id": "m202",
      "content": "For internal review notes, I prefer brief text with evidence links because I read them between meetings.",
      "reasoning": "Both entries describe the same internal-note preference and reason."
    },
    {
      "kind": "archive",
      "memory_id": "m201",
      "reasoning": "Same preference and reason are preserved in the proposed m202 replacement."
    }
  ]
}

A misleading proposal set to reject

{
  "proposals": [
    {
      "kind": "replace",
      "memory_id": "m202",
      "content": "Keep all my writing short.",
      "reasoning": "Simplified preference."
    },
    {
      "kind": "archive",
      "memory_id": "m201",
      "reasoning": "Duplicate."
    }
  ]
}

Broadens internal notes to all writing and drops both evidence links and the reason.

3. A narrow opt-out stays narrow

Synthetic input memories

• m302, created 2026-10-29: Stop including Orchid Fund in the weekly research digest because that team opted out. Keep other funds on the digest.
• m301, created 2026-10-20: Prepare a weekly research digest for Orchid Fund and Birch Fund.

One acceptable proposal set

{
  "proposals": [
    {
      "kind": "replace",
      "memory_id": "m302",
      "content": "Prepare the weekly research digest for Birch Fund. Exclude Orchid Fund because that team opted out; other funds remain on the digest.",
      "reasoning": "Combines the continuing instruction with the later Orchid-only opt-out."
    },
    {
      "kind": "archive",
      "memory_id": "m301",
      "reasoning": "The proposed m302 replacement preserves the continuing scope and later correction."
    }
  ]
}

A misleading proposal set to reject

{
  "proposals": [
    {
      "kind": "replace",
      "memory_id": "m302",
      "content": "Stop the weekly research digest for all funds.",
      "reasoning": "Applied latest stop instruction."
    },
    {
      "kind": "archive",
      "memory_id": "m301",
      "reasoning": "Superseded."
    }
  ]
}

Turns a narrow revocation into a global stop and loses why it was issued.

All records are invented. Equivalent wording is fine. These illustrate full proposal sets; replace/archive suggestions remain separately reviewed in the UI. They are reviewer examples, not observed failures or executed tests.