Claude Opus 4.7 Multi-Session Memory M&A Diligence

Claude Opus 4.7 multi-session memory is the feature M&A practices have been waiting for since 2023. Anthropic shipped it on April 16, 2026, alongside task budgets and the new "xhigh" effort level. Per Anthropic's What's New in Claude Opus 4.7 docs, Claude can now persist context across sessions via a scratchpad/notes file. For a 12-day diligence engagement, that means the model resumes where it left off — same parties, same facts, same line of analysis — instead of being re-introduced every morning. The recovered analyst hours per matter are real money. The data-handling implications are also real, and most M&A AI policies haven't caught up.

What multi-session memory actually is

Claude 4.6 sessions started cold. You loaded the matter, explained the parties, grounded the model in the facts, and started the analysis. Tomorrow morning, you did it again. That re-priming step was the *context-loss tax* — the operational reason AI was useful for one-shot tasks and frustrating for long-horizon work.

Opus 4.7 ships scratchpad/notes file persistence by default. Claude writes structured notes mid-session capturing the matter context, party identities, key facts, and analytical pathways already explored. You save the file with the matter. The next session reads it back. The model resumes mid-analysis instead of restarting.

The scratchpad is a real file on your storage; not a server-side memory. Anthropic doesn't retain it; the firm does. That makes it a firm-data asset from session one, with all the document retention, access control, and confidentiality implications that follow. The Opus 4.7 anchor covers the broader change set.

Why M&A diligence specifically benefits

M&A diligence runs 5-15 days. The work involves repeated context loading: parties, deal structure, term sheet, prior negotiations, jurisdiction, regulatory posture, sector-specific risk. Every analyst session re-grounds in the same baseline before getting to new work.

On Opus 4.6, that re-grounding consumed 30-60 minutes of analyst time per session, plus token spend on re-loading documents the model had already processed. For a 12-day engagement with 2-3 sessions per day per analyst, that's 12-25 hours of redundant context loading per analyst. At a $400/hour blended rate, the recovered hours add up to $4,800-$10,000 per matter.

With multi-session memory, the analyst opens the scratchpad, Claude reads the context summary, and work resumes inside 5 minutes. The recovered hours stay billable to substantive analysis instead of being absorbed by overhead.

The second-order effect: M&A practices running 6-10 simultaneous diligence engagements compound the recovered time across the portfolio. A practice doing 8 active diligences saves 100-200 analyst hours per month; the equivalent of half a junior associate's capacity, recovered. The third-order effect: that recovered capacity can absorb additional matters without growing headcount, which changes the partner-track economics of mid-market M&A practices. The task budgets discovery spoke covers the parallel cost-side feature.

How to structure a scratchpad for an M&A engagement

The scratchpad isn't free-form. It's a structured notes file Claude reads on session restart. A practical M&A scratchpad has six sections:

Deal context; parties, target, acquirer, deal type, indicative consideration, signing target date.

Party identity register; full legal names, jurisdictions of incorporation, key entities in the corporate family, KYC status for sanctioned-party screening.

Key documents loaded; term sheet version, draft purchase agreement version, data-room index date, third-party diligence reports referenced.

Analysis state; areas already covered (e.g., "reps and warranties scrubbed for material qualifiers"), areas in progress, areas not yet started.

Open questions register; questions raised by the model, partner, or client that haven't been resolved with a documented answer yet.

Risk flags; issues escalated to partner attention, with cross-references to source documents.

Claude maintains this scratchpad across sessions. You can audit it, edit it, version-control it, and store it with the matter file. For multi-analyst teams, the scratchpad becomes the shared coordination layer; same model, same context, same risk framing across the diligence team.

Storage, retention, and confidentiality

Scratchpad files contain matter-specific reasoning, party identities, and analysis pathways. They're privileged work product when generated under attorney direction, and they're confidential client information regardless. That makes storage policy a Day 1 question.

Three storage models that fit firm data-governance norms:

Matter folder integration. Store the scratchpad in the same matter folder as drafts, due diligence reports, and client correspondence. The retention schedule, access control, and audit trail inherit from existing matter management infrastructure. This is the cleanest fit for firms with established matter management systems.

Practice-area shared workspace. For multi-analyst teams, store the scratchpad in a shared practice-area workspace with role-based access. Useful for handoffs across associates and for partner review.

Encrypted firm-wide storage with retention rules. For firms running formal information governance programs, encrypted storage with defined retention windows and litigation-hold capability. Adds operational overhead but gives the cleanest defense in a discovery dispute.

The AI policy must address: who can create scratchpads, where they get stored, how long they're retained post-matter, and what the destruction protocol is. The cybersecurity safeguards privileged context spoke covers a parallel policy update for confidentiality.

Privilege implications: when does the scratchpad become discoverable

The scratchpad raises real privilege questions that 4.6 didn't. On 4.6, Claude's session ended with the conversation; nothing persisted. On 4.7, the scratchpad is a document. Documents are discoverable if not privileged.

United States v. Heppner (SDNY, Feb 17, 2026) ruled that defendant Bradley Heppner's exchanges with consumer Claude weren't protected by attorney-client privilege or work-product doctrine because Claude isn't an attorney and Heppner generated the materials independently of counsel direction. The Heppner ruling addressed consumer Claude. Enterprise scratchpads generated under attorney direction sit in a different posture; they're created at counsel's direction in anticipation of (or during) litigation or a transaction. That fits the work-product framework better than Heppner did.

But it's not automatic. The scratchpad must be (a) generated under attorney direction, (b) reflect attorney mental impressions and analysis, and (c) be maintained as confidential work product. Practical implications:

- Document the matter team's involvement in scratchpad creation in the engagement letter or matter intake. - Treat scratchpads as work product in document holds and discovery responses. - Don't share scratchpads with non-counsel parties; that may waive work-product protection.

The Heppner explainer covers the underlying ruling. The work-product analysis for scratchpads is jurisdiction-specific; coordinate with the firm's risk-and-ethics committee before establishing default practice.

What multi-session memory doesn't fix

Three operational realities that persist even with scratchpad persistence:

The model still hallucinates without verification. Multi-session memory means Claude remembers what it told you yesterday. It doesn't mean what it told you yesterday was right. Citation verification against Westlaw, Lexis, or primary sources remains mandatory for legal claims. Per Anthropic's documentation, calibration improvements in 4.7 reduce overconfident hallucinations but don't eliminate them.

The scratchpad has token costs. Reading the scratchpad on session restart consumes input tokens. For long matters, the scratchpad grows. Plan periodic compaction; let Claude summarize the scratchpad and replace verbose history with the summary. The tokenizer cost calculator covers the input-token math.

Multi-analyst handoffs still need explicit coordination. The scratchpad captures Claude's view of the matter, not the human team's. Use the scratchpad as the model context layer; use existing matter management tools (matter notes, partner instructions, working group calls) for human coordination. Don't conflate the two layers.

For mid-market M&A practices that don't run dedicated information governance programs, the practical approach is to treat scratchpads like any other matter note: matter-folder storage, standard retention, work-product handling on discovery.

The Bottom Line: The verdict: multi-session memory turns Claude from a one-shot tool into a long-horizon analyst layer. M&A practices recover 12-25 analyst hours per diligence matter. Mid-market practices running 6-10 simultaneous diligences gain effective half-associate capacity. The procurement math is straightforward; Claude Team at $25 per user per month pays back in week one for any active M&A practice. The harder work is the storage and privilege policy, which has to ship before the first scratchpad does.