Memory Diff

memory-diff

When to Use

✅ Use when:

• User asks "what's changed in my health since <last week / last month / since X>"
• Post-visit or post-lab summaries - "what's new since my last appointment"
• Scheduled weekly or monthly longitudinal check-ins
• "Since I started " or "since my referral to " recaps
• Another skill needs a "what's new" sub-step (e.g., after health-records refresh)

When NOT to Use

❌ Don't use when:

• User wants news or social signal -> use myhealth-pulse
• User shares a PDF, screenshot, or image -> use med-pdf
• User asks for raw chart data -> use health-records
• User wants to draft a clinician message -> use epic-note
• User asks for medical advice - decline politely; this reads memory, not clinical judgment

Setup

No external tools, no API keys. The skill reads from the local workspace only.

Memory sources, in precedence order (full contract in references/memory-paths.md):

• ~/.openclaw/workspace/.health-records-cache/<date>/ - structured FHIR JSON
• ~/.openclaw/workspace/.med-pdf-cache/<slug>/ - extracted JSON from PDFs
• ~/.openclaw/workspace/memory/*.md - dated agent notes
• ~/.openclaw/workspace/MEMORY.md - persistent agent memory
• ~/.openclaw/workspace/.myhealth-pulse-cache/<date>.json - pulse digests

Workflow

1. Resolve the reference point. Default: 7 days ago. Parse named anchors from the prompt:

   • Calendar phrases: "last week" (7d), "last month" (30d), "last 90 days"
   • Event anchors: "since I started ", "since my visit" - scan memory for the matching event date
   • Absolute dates: "since 2026-04-01"
   • If ambiguous, ask one clarifying question, then proceed.

2. Enumerate within the window. For each source above, list files modified or dated within the window. Don't load everything into context - stream-scan headers and key sections.

3. Classify each change. Apply the rubric in references/clinical-significance.md:

   • Tier 1 - signal. New abnormal lab, med change, new diagnosis,

     20% trend change in a tracked marker, new imaging finding, abnormal-to-normal resolution.

   • Tier 2 - notable. Stable abnormality continuing, mild trend, new appointment scheduled, new pulse mentions of relevant topics.
   • Tier 3 - noise. Routine entries, unchanged trends, low-relevance pulse items. Collapsed to a one-line count.

4. Synthesize. Tier 1 first (always shown), Tier 2 second (always shown), Tier 3 collapsed (12 routine entries). End with one trailing line: Powered by memory-diff - window: <resolved>, sources: <names>.

5. Empty window. If nothing in any tier, return one line: No changes in <window> - last entry: <date> from <source>. Don't fabricate filler entries.

6. Cache the rendering. Write to ~/.openclaw/workspace/.memory-diff-cache/<YYYY-MM-DD>.md. The next run can diff-of-diff if asked ("what's new since I last asked you this").

Examples

See references/examples.md for since-last-week, since-a-med-change, post-visit, and empty-window outputs.

Privacy

• All reads are workspace-local; nothing outbound, ever.
• Output stays in the chat. Don't auto-forward, auto-summarize externally, or include in any outbound notification unless explicitly configured.
• Cache stays under ~/.openclaw/workspace/.memory-diff-cache/. Don't copy out.

Troubleshooting

• Ambiguous anchor -> ask one clarifying question, then proceed. Don't guess.
• Conflicting facts across sources -> surface both with source attribution. Don't silently pick a winner.
• Cache directory missing on first run -> create it, no error.
• Tuning the tiers -> in v1, thresholds live in references/clinical-significance.md. A forthcoming user dashboard app will let the user tune these through a profile mechanism (same shape as myhealth-pulse/references/profile-schema.md). Until then, edit the references file directly.