Cursor For Accountants
Typed engagement memory AI agents can actually think in.
Every accounting firm is deploying AI agents. They all break the same way: by file 4 the context is full, by file 8 the model is hallucinating numbers. We built an opinionated financial schema the agent is forced to write its findings into. Typed entities, FSLI-level reconciliations, period-linked deltas, invariant-checked, every fact traced back to the source cell. Engagement Memory is the LSP for financial work.
Typed entities, not free-text
The agent is forced to record findings as typed objects: vendors, accounts, periods, line items. No scratchpad, no hallucinated structure.
FSLI-level reconciliation
Every number lives in a Financial Statement Line Item with a canonical mapping. Cross-document deltas surface automatically.
Period-linked deltas
Q3 2025 revenue is reconciled against Q3 2024 the moment the file lands. Period boundaries are typed, not parsed from headers.
Provenance to source cell
Every finding traces back to a specific cell in a specific sheet in a specific file. Reviewers verify in one click.
Engagement Memory
How the schema turns every uploaded file into structured intelligence.
Engagement Memory does three things. It catches discrepancies the moment a file lands, because the reconciliation runs on every write. It creates the structured data layer AI agents need to reason across multiple files at once. And for recurring engagements, the memory persists across periods so numbers stay consistent with what was previously reported.
A document lands. The schema receives it.
The system parses the workbook, identifies the document type, and writes typed facts into Engagement Memory. The agent never touches free-form text. Every fact is structured before it counts.
Every fact reconciles against everything already in the engagement.
New deposits tie to existing revenue rows. New trial balances cross-check against prior periods. Accounting invariants run on every write. The schema enforces the relationships.
Findings surface inline, with provenance.
Discrepancies appear in-cell with click-through to the source file, sheet, and cell. The reviewer accepts or rejects in place. The memory deepens with every accepted finding.
The LSP analogy
Cursor is not valuable because of the AI. It is valuable because the agent reads and writes through a typed language server. Engagement Memory is the LSP for financial work.
Correctness By Construction
The schema is the moat. Correctness is the proof.
Generic agent memory tools are vector stores and scratchpads: free-form text the agent dumps into and retrieves from. Engagement Memory is the opposite. The schema is enforced by accounting invariants, every write is reproducible, and every fact has a pointer back to the cell it came from.
Deterministic correctness, not statistical guessing
The schema is enforced by accounting invariants: debits equal credits, balance sheet ties, period coverage holds. Same inputs produce the same findings every run, and every finding is reproducible on demand.
Provenance you can stake an audit on
Every typed fact in Engagement Memory carries a pointer to its source file, sheet, and cell. Findings link back the way LSP go-to-definition links to source code.
Customer uploads are not training fuel
Customer files support the current engagement workflow. They are not used to train shared models, and the page says that plainly.
Live now
The schema is shipped. Pilot teams write into it today.
- Typed Engagement Memory writes with FSLI canonical
- Cross-document reconciliation on every new fact
- Period-linked deltas across prior and current periods
- Accounting invariants enforced on every write
- Provenance to source file, sheet, and cell
- Tenant-scoped data room ingest with audit trail
Coming next
Auto-ingestion is the next milestone, not today's claim.
Documents classified in V1 move into the normalized ingestion path automatically. "You received 31 files. 28 ingested. 3 need review." That is the next layer of the product, not the thing pilot teams are being asked to trust today.
Customer Category
Built for cross-party financial work.
Any engagement where one party is reviewing, analyzing, or reconciling financial documents provided by another party. That is QoE, FDD, audit, tax preparation, 409A valuation, fund administration, credit underwriting, forensic accounting, and every professional service where a firm is making sense of a counterparty's books. We are not a close tool. We are not an ERP. The structural relationship is why the product has a data room, request lists, and engagement memory.
Cross-party (us)
- QoE / FDD advisory
- External audit
- Tax preparation
- 409A valuation, ASC 805
- Fund administration
- Credit underwriting
- PE portfolio monitoring
- Forensic accounting
- Restructuring advisory
Intra-party (not us)
- Internal month-end close
- Bookkeeping for own company
- FP&A and budget reporting
- AP / AR operations
- Payroll
- Treasury management
- ERP migration (own data)
- Internal controls testing
Different category, different buyer. The intra-party tools (Basis, Campfire, Rillet, Numeric, FloQast) serve a company closing its own books. SynthGL serves the firm across the table reviewing those books.
Narrow Gtm, Wide Category
Start with QoE / FDD. The category goes much further.
The initial buyer is a QoE / FDD team where the pain density is highest and the workflow shape maps cleanly to Engagement Memory. The expansion path matters, but it stays subordinate to the first workflow we are actually trying to win.
Founder Credibility
An accountant who got tired of the workflow and built the data layer to fix it.
The QoE workflow lived end to end at PwC Deals and at a boutique FDD shop, then a wall: scripts that worked on one client's files broke on the next. SynthGL is what came out of getting frustrated enough to build the typed data layer underneath.
PwC CMAAS, then boutique FDD
Three years at PwC Deals on technical accounting for IPOs, M&A, and carve-outs, including the largest 2023 IPO. Then six months at a boutique FDD shop running the exact Quality of Earnings workflow Engagement Memory now automates.
The schema is the moat, and it already exists
Behind the pilot is a typed financial data platform: FSLI canonical, entity resolution, period model, reconciliation taxonomy, and accounting invariants. The components a horizontal AI memory tool would have to rebuild from scratch.
Honest scope, not AI theater
The page is explicit about what is live (typed engagement memory, reconciliation, provenance, audit trail) and what is queued (auto-ingestion, V1.5). One layer shipped honestly beats the whole stack promised on day one.
FAQ
The questions investors and pilot teams actually ask.
A managing director, an investor, and an analyst should each be able to land here and find the answer that matters to them without hunting for the trust story between the lines.
Pilot Program
One CTA. One next step.
No pricing cards, no demo / booking / waitlist split. If your engagement looks like cross-party financial work, apply for the pilot and we will review fit.
Best fit for the first pilots
QoE / FDD teams running structured request lists across PE deal volume. The workflow shape (trickle-in client files, recurring cadence, cross-document reconciliation) maps directly onto Engagement Memory.
What pilot teams evaluate
Whether the typed schema captures the findings their analysts would have caught manually, whether the cross-document reconciliation catches things faster, and whether the provenance trail is strong enough for real deal work.
What happens after you submit
We review fit, follow up on your workflow, and use the pilot to shape the next hardening pass. The commercial model is intentionally not locked yet.
Apply for the pilot.
Tell us about the engagement type, the workflow today, and where the cross-document reconciliation work slows the team down.