Cross-Deal Intelligence

Every memo you generate makes the next one a little sharper. Memosa does not analyze each deal in isolation — it builds a structured understanding of the properties, sponsors, and markets across your team’s portfolio and feeds that back into new analyses. This page explains what that means for you as an analyst: what cross-deal intelligence does, where it shows up in Canvas, and what it deliberately does not do.

Note

Cross-deal intelligence is scoped to your team. Patterns, comparables, and sponsor history are visible only within your team’s deals. Another team’s identically-named sponsor never collides with yours — every cross-deal record is keyed by team. See Roles and Permissions for how access works at the deal level.

The four kinds of cross-deal knowledge

Memosa reuses knowledge across deals in four distinct ways. Each is independent — you benefit from all of them without configuring anything.

1. Comparables (the property graph)

When you upload a deal’s documents, Memosa extracts the subject property’s attributes — asset class, size, vintage, submarket — and records them in a per-deal property graph. When the comparables section is written, it queries that graph with typed real-estate constraints rather than fuzzy text similarity:

• Same compatibility group — multifamily compares to multifamily, office to office, and so on.
• Size band — within ±50% of the subject’s size .
• Vintage band — within ±10 years of the subject’s year built .
• Submarket (optional) — same submarket when you ask for it.

This typed approach excludes “same-class but out-of-band” comps that a similarity score would otherwise surface — a 400-unit garden complex no longer shows up as a comp for a 40-unit infill deal just because both are multifamily.

Where you see it: the comparables section of the memo, plus a Comp filters tab in the Canvas right panel that shows the exact constraints the comparables agent operated on — the compatibility group, the active size and vintage bands, and the ranked comps that passed.

2. The CRE ontology (one vocabulary, everywhere)

Behind the scenes, every part of Memosa shares a single source of truth for real-estate taxonomy: what counts as an asset type, a deal type, a memo section, a comp compatibility group, and how data sources rank against each other. A metric called “NOI” means the same thing in the financial model parser, in retrieval, and in the synthesis prompt — there is no drift between subsystems.

You never edit the ontology, but you feel its effects: consistent section structure, comps grouped the way a real underwriter would group them, and conflict resolution that always applies the same precedence rules (described below). For the engineering view of why this matters — and the production cascade that motivated it — see the architecture page on the system overview.

3. The playbook (learned patterns)

As you and your team edit, rate, and approve memos, Memosa learns which analytical approaches hold up. These are stored as a playbook of patterns, promoted through tiers as evidence accumulates and decayed when they go stale. The playbook informs how deep-research agents approach a new section — not by copying text, but by recalling the shape of analysis that worked before.

Two guardrails keep the playbook safe and useful:

• No deal-specific data leaks in. A strict content filter rejects any candidate pattern that contains dollar amounts, financial ratios, addresses, entity names, or specific lease terms. Patterns capture approach, never another deal’s numbers.
• Asset classes stay separate. Multifamily patterns are namespaced away from office and retail, so a residential playbook never contaminates a commercial analysis.

4. Sponsor patterns (cross-deal sponsor history)

When the same sponsor appears across multiple deals, Memosa recognizes them as one entity even when the name is spelled differently from deal to deal (“Acme Capital LLC” and “Acme Capital, L.P.” resolve to the same canonical sponsor). Once a sponsor is tied to two or more deals , their history surfaces as a cross-deal pattern in the sponsor-background section.

Where you see it: a Sponsor patterns tab in the Canvas right panel. Cross-deal sponsors are flagged with a badge showing how many of your deals they appear in; single-deal observations are shown separately.

Tip

Cross-deal sponsor recognition is probabilistic and conservative by design. Memosa merges two sponsor names only when it is highly confident they are the same entity — it would rather treat two genuinely-different sponsors separately than incorrectly fuse them, because a wrong merge would pollute the sponsor history for every deal that sponsor touches.

How data sources are reconciled across a deal

When two documents disagree about a number — say the Excel model and a PDF both state a cap rate — Memosa does not guess. It applies a fixed precedence order, highest wins:

Source	Precedence	Typical role
Your direct input	100	Corrections and figures you type in chat
Excel model	90	Pro-forma, cash flow, returns, debt schedule
CoStar export	70	Market, demographics, comparables
PDF (OM, etc.)	50	Narrative, sponsor, property description

This is why a figure in the memo can differ from one in your PDF: a higher-precedence source (usually the Excel model) supplied the authoritative value. When sources conflict, the memo footnotes the resolution so you can see exactly which source won and why.

Where you see it: the Provenance tab in the Canvas right panel shows, per resolved conflict, the winning source type, its weight, and the candidates that were considered. If a memo number surprises you, this is the first place to look — see also Retrieval and Evidence.

What cross-deal intelligence does not do

Caution

• It does not copy text, numbers, or conclusions from one deal into another. Comps reference real attributes; patterns capture approach; sponsor history is observational. The actual analysis is always grounded in this deal’s documents.
• It does not cross team boundaries. There is no shared global pool — your portfolio is yours.
• It is not a substitute for evidence. If a deal is missing a CoStar export, cross-deal knowledge cannot fabricate market data. Upload the right document type for the section you care about (Supported Documents).

Going deeper

• Portfolio Insights — aggregate trends across your finalized deals (with privacy thresholds).
• Quality and Readiness — how Memosa scores a memo and learns from your edits.
• Retrieval Pipeline and Reranking — the engineering view of how evidence is gathered and ranked.

Sources

• src/intel/ontology/cre_ontology.yaml — CRE taxonomy and precedence_weights (USER_INPUT/EXCEL/COSTAR/PDF).
• src/intel/deal_graph/ — per-deal property graph, entities, relations, claims (store.py, traversal.py, population.py).
• src/canvas/routes/comparables_trail_routes.py — comp constraint defaults surfaced to the UI (_DEFAULT_SIZE_BAND_PCT, _DEFAULT_VINTAGE_BAND_YEARS).
• src/intel/deal_graph/cross_deal_store.py — cross-deal sponsor entity store; min_deal_count threshold for surfacing patterns.
• src/canvas/routes/sponsor_patterns_routes.py, src/canvas/routes/conflict_provenance_routes.py — the right-panel data surfaces.
• src/canvas/auth/permissions.py — ROLE_HIERARCHY (team/deal access scoping).
• Native memory: cre_ontology.md (ontology + property graph + cross-deal entity resolution, Tiers 1–6), playbook_memory_system.md (tier promotion, confidence decay, DealContentFilter rules, asset-class namespacing).