Turn a Harvey Memory Dump into a Client-Matter Privilege Log in One Prompt Session

The task

An associate or knowledge-management attorney exports their Harvey interaction history — now persistent across matters thanks to Harvey II's Memory capability, alongside preference setting, context sharing, and learning from both what a law firm's lawyers do — and needs to hand a clean, matter-scoped privilege log to the responsible partner before a document production or an internal audit. This comes up whenever you switch off a matter, respond to a discovery request, or route work-product to co-counsel.

Before AI

You'd open the export (JSON or CSV), skim each interaction, tag it by matter, decide whether it reflects legal advice or attorney work-product, and hand-type entries into a Word or Excel privilege log using firm format. For a mid-sized matter (50-150 interactions) this is 60-120 minutes of tedious classification, and the log quality depends entirely on how alert you are at hour two.

The workflow

Harvey's new Memory layer is useful precisely because it can retain drafting habits, stylistic preferences and matter-specific context so that future work reflects prior decisions without requiring lawyers to restate instructions or search for precedent. But that same persistence means the raw log now mixes matters, and each entry is potentially privileged communication or work-product that has to be tracked as such. Below is a three-step prompt session that turns the export into a defensible draft log.

Step 1 — Classify and matter-tag each interaction

Paste the raw export below the prompt. The model reads it as text; no file handling required.

You are a senior litigation associate preparing a privilege log from a raw AI-assistant interaction export. The export is appended below this instruction block as JSONL or CSV-like text; each row is one interaction (with timestamp, user, prompt, and response snippet).

For EACH row, output a JSON object with these fields:
- row_id: use the "id" field if present, else a 1-based index
- timestamp: ISO date
- author: the attorney's name
- matter_guess: the client-matter this most likely belongs to, inferred from party names, deal codes, or subject cues in the prompt/response. If unclear, output "UNASSIGNED".
- privilege_basis: one of ["Attorney-Client","Work Product","Both","None"]
- privilege_rationale: one sentence explaining why (or why not) — reference the specific text cue.
- description: a neutral, non-waiving description suitable for a privilege log (e.g., "Analysis prepared at attorney direction regarding indemnity carve-outs"). Do NOT quote the substantive content.
- redaction_flag: true if the row itself contains client confidences that should not appear verbatim in the log.

Return a single JSON array. No prose before or after. If a row is ambiguous, still emit a record and set privilege_basis to your best call with rationale.

id,timestamp,author,prompt,response_snippet
H-1041,2026-08-14T09:12:00Z,A. Ramirez,"Draft an indemnity carve-out for Project Larkspur MSA limiting seller liability for third-party IP claims to actual damages","Suggested language: 'Notwithstanding Section 9.2, Seller's liability for Third-Party IP Claims shall be limited to direct damages actually paid...'"
H-1042,2026-08-14T09:47:00Z,A. Ramirez,"Compare this redline to our last Larkspur turn - is Buyer pushing on the basket?","Yes - Buyer raised the basket from $250k to $400k and added a mini-basket for reps 3.4-3.7."
H-1043,2026-08-14T14:03:00Z,A. Ramirez,"What's the boilerplate NY choice-of-law clause we use for SaaS deals","Standard clause: 'This Agreement shall be governed by and construed in accordance with the laws of the State of New York...'"
H-1044,2026-08-15T08:22:00Z,A. Ramirez,"Summarize privileged strategy memo re Meridian Foods class cert opposition - focus on Daubert angles","Three Daubert angles were flagged: (1) Dr. Chen's damages model lacks class-wide methodology; (2) survey sample size..."
H-1045,2026-08-15T10:15:00Z,A. Ramirez,"draft email to opposing counsel on Meridian scheduling","Draft: 'Counsel - Further to our meet-and-confer, we propose the following deposition schedule...'"
H-1046,2026-08-15T11:40:00Z,A. Ramirez,"Larkspur - client asked whether we can accept a 24-month survival on fundamental reps. What's our fallback?","Fallback position: accept 24 months only if paired with an uncapped indemnity for fundamentals and a tightened definition..."
H-1047,2026-08-15T15:30:00Z,A. Ramirez,"pull the standard force majeure language","Standard FM clause per firm precedent bank..."
H-1048,2026-08-16T07:55:00Z,A. Ramirez,"Meridian - strategy for deposing Dr. Chen next Tuesday, focus on methodology gaps","Suggested outline: (1) qualifications challenge on class-wide extrapolation; (2) prior inconsistent testimony in Hartwell v..."

Step 2 — Roll up to a per-matter log with a privileged/non-privileged split

Take the JSON array you just produced and reshape it into a privilege log grouped by matter_guess.

For each matter, output:
1. A markdown H3 heading: "### Matter: <name>"
2. A markdown table with columns: Log No. | Date | Author | Description | Privilege Basis | Notes
   - Log No. format: <MatterInitials>-PL-001, incrementing per matter.
   - Include ONLY rows where privilege_basis is Attorney-Client, Work Product, or Both.
   - Description must be the non-waiving description from step 1.
   - Notes column should flag redaction_flag=true rows with "REDACT — contains client confidences" and any UNASSIGNED matter guesses with "CONFIRM MATTER".
3. After the table, a bullet list titled "Excluded as non-privileged (for reviewer confirmation):" listing row_ids with privilege_basis = None and a one-line reason.

At the very end, output a "## Reviewer checklist" section with:
- Count of entries per matter
- Any UNASSIGNED rows the partner must resolve before finalizing
- Any rows where privilege_basis was "Both" (these often warrant closer review)

Preserve the classifications from step 1 exactly — do not re-classify.

Step 3 — Draft the partner cover memo

Now draft a short cover memo (200-300 words) from the associate to the responsible partner, transmitting the privilege log above. Format:

TO: [Responsible Partner]
FROM: [Associate]
DATE: [today]
RE: Draft Privilege Log — AI-Assistant Interactions, [list matter names from the log]

The memo should:
- State the source (Harvey Memory export) and date range covered (infer from the log).
- Note that all entries reflect either legal advice sought/given or work-product prepared in anticipation of litigation or in connection with client representation.
- Explicitly flag any UNASSIGNED rows and any "Both" privilege entries as needing partner confirmation.
- Note that non-privileged rows (boilerplate/precedent lookups) were excluded and are listed for the partner's verification.
- Close with a request for sign-off before the log is entered into the firm's matter management system.

Keep the tone professional and hedged. Do not overstate the completeness of the classification — this is a first-pass draft prepared by an associate for partner review.

Gotchas

• Matter inference is a guess, not a determination. The model infers matter from party/deal-code cues. If your firm uses opaque matter numbers (e.g., "48221-003"), the matter_guess field will mostly return UNASSIGNED. Add a matter-code glossary to Step 1 if you want better precision.
• "Boilerplate lookup" isn't automatically non-privileged. A generic force-majeure query in isolation may look non-privileged, but if it's tied to a specific matter, the request itself can reveal privileged strategy. The Step 1 rationale field is where you audit this — read it before accepting the exclusions.
• Never paste the export raw into a public model. Use your firm's approved enterprise deployment. The prompts above are model-agnostic but the data isn't.
• Descriptions can waive privilege if too specific. The prompt tells the model to write non-waiving descriptions, but the partner should still spot-check every row before the log leaves the firm.
• DeepL translation adds a wrinkle. Since Harvey brings DeepL's specialised Language AI capabilities to legal teams around the world, who work across borders, cross-border matters may have interactions in multiple languages. Add "translate non-English rows to English before classifying, and preserve the original in a source_text field" to Step 1 if needed.
• Harvey II's Memory is designed to persist across matters — the new Harvey platform will support user preferences, matter history and context with the help of Memory — which is exactly why this log needs to be run regularly, not just at matter close.

Time saved

Estimate: 45-90 minutes per matter export of 50-150 interactions, versus manual classification and log-typing. The savings come from (a) skipping the row-by-row read-and-tag pass, and (b) auto-generating the non-waiving descriptions and log numbering. This is an estimate based on typical mid-matter export sizes; complex litigation exports with hundreds of interactions across many matters will save proportionally more, but also demand more partner review time on the back end. Treat the output as a first-pass draft for attorney review, never a filing-ready document.

What we got when we ran it

We executed every prompt above, in order, against the sample input (model: claude-sonnet-4-6). Outputs below are real and unedited apart from length truncation.

Step 1 output

```json [ { "row_id": "H-1041", "timestamp": "2026-08-14", "author": "A. Ramirez", "matter_guess": "Project Larkspur – MSA Negotiation", "privilege_basis": "Both", "privilege_rationale": "The prompt requests drafting of a specific contractual provision for a named client matter ('Project Larkspur MSA'), reflecting confidential client instructions communicated to counsel, and the resulting draft constitutes attorney work product prepared in anticipation of or during a deal.", "description": "Attorney-directed drafting of indemnity carve-out language for master services agreement, addressing limitation of seller liability for third-party intellectual property claims.", "redaction_flag": true }, { "row_id": "H-1042", "timestamp": "2026-08-14", "author": "A. Ramirez", "matter_guess": "Project Larkspur – MSA Negotiation", "privilege_basis": "Both", "privilege_rationale": "The prompt involves attorney analysis of a counterparty redline for a named client matter, and the response reflects counsel's legal assessment of negotiating positions, constituting both attorney-client communication and opinion work product.", "description": "Attorney analysis of counterparty redline to transaction agreement, assessing changes to indemnification basket provisions across identified representation categories.", "redaction_flag": true }, { "row_id": "H-1043", "timestamp": "2026-08-14", "author": "A. Ramirez", "matter_guess": "UNASSIGNED", "privilege_basis": "None", "privilege_rationale": "The prompt requests retrieval of standard boilerplate firm precedent language with no client-specific facts or confidential communications, and the response is a form clause from a general precedent bank rather than legal advice rendered to a specific client.", "description": "Retrieval of firm standard-form New York choice-of-law clause for SaaS agreements from precedent bank.", "redaction_flag": false }, { "row_id": "H-1044", "timestamp": "2026-08-15", "author": "A. Ramirez", "matter_guess": "Meridian Foods – Class Certification Opposition", "privilege_basis": "Both", "privilege_rationale": "The prompt expressly references a 'privileged strategy memo' and

…output truncated at 250 words.

Step 2 output

Matter: Project Larkspur – MSA Negotiation | Log No. | Date | Author | Description | Privilege Basis | Notes | |---|---|---|---|---|---| | LP-PL-001 | 2026-08-14 | A. Ramirez | Attorney-directed drafting of indemnity carve-out language for master services agreement, addressing limitation of seller liability for third-party intellectual property claims. | Both | REDACT — contains client confidences | | LP-PL-002 | 2026-08-14 | A. Ramirez | Attorney analysis of counterparty redline to transaction agreement, assessing changes to indemnification basket provisions across identified representation categories. | Both | REDACT — contains client confidences | | LP-PL-003 | 2026-08-15 | A. Ramirez | Attorney analysis of client-directed inquiry regarding survival period for fundamental representations, including identification of fallback negotiating position and related indemnification conditions. | Both | REDACT — contains client confidences | Excluded as non-privileged (for reviewer confirmation): - H-1043 — Retrieval of standard boilerplate New York choice-of-law clause from firm precedent bank; no client-specific facts or confidential communications present. - H-1047 — Retrieval of standard form force majeure clause from firm precedent bank; no client-specific facts, legal analysis, or confidential communications present. --- ### Matter: Meridian Foods – Class Certification Opposition | Log No. | Date | Author | Description | Privilege Basis | Notes | |---|---|---|---|---|---| | MF-PL-001 | 2026-08-15 | A. Ramirez | Summary of privileged litigation strategy memorandum prepared in connection with class certification opposition, addressing anticipated Daubert challenges to opposing expert's methodology and opinions. | Both | REDACT — contains client confidences | |

…output truncated at 250 words.

Step 3 output

--- TO: Responsible Partner FROM: A. Ramirez DATE: 2025-07-14 RE: Draft Privilege Log — AI-Assistant Interactions, Project Larkspur – MSA Negotiation; Meridian Foods – Class Certification Opposition --- Please find attached a first-pass privilege log prepared from an export of Harvey Memory AI-assistant interaction records covering the period August 14–16, 2026. The log addresses interactions across two matters: Project Larkspur – MSA Negotiation and Meridian Foods – Class Certification Opposition. Each logged entry reflects either (i) a request for, or communication of, legal advice in the context of an attorney-client relationship, or (ii) material prepared in anticipation of litigation or in connection with client representation constituting protected work product. Privilege designations are based on a review of prompt and response content and should be treated as a preliminary assessment only, subject to your review and approval. A few items warrant your specific attention before the log is finalized: - UNASSIGNED rows (H-1043, H-1047): Two interactions — both involving retrieval of standard firm precedent clauses — could not be tied to a confirmed client matter. These were classified as non-privileged on the current record, but if either was retrieved in connection with a specific engagement, the matter assignment and privilege basis may require revision. - "Both" designations (H-1041, H-1042, H-1044, H-1046): Four entries carry a dual Attorney-Client/Work Product classification. Given that courts occasionally scrutinize dual-basis assertions, I would appreciate your confirmation that the descriptions are sufficiently neutral and that the designations are supportable on the facts. - **Non-privileged exclusions (H-1043, H-1045,

…output truncated at 250 words.

---

This content is for informational purposes only and is not legal advice. Confirm confidentiality, privilege, and jurisdictional rules before using any AI tool with client matters.