Storage architecture for crypto funds: where on-chain, off-chain, and AI logs actually live

Crypto fund storage is a three-tier problem dressed as a one-tier problem

The default approach is: "I have my transaction history in a Google Sheet, my contracts on Notion, and my wallet addresses in a 1Password vault." That's not storage architecture. That's a series of accidents waiting to compound.

The right architecture has three tiers, each with different durability, access, and audit requirements.

Tier 1 — Hot reconciliation (working data, daily access)

Lives on: workstation NVMe + cloud-synced (Dropbox Business, Box, or S3-backed).

Contains: - Current month's close working files - Wallet API pulls for the period being reconciled - Open exception lists for the AI agents to resolve - In-flight Xero/QuickBooks ledger exports

Retention: rolling 90 days, then move to Tier 2.

Why this matters: the AI agents need fast, deterministic access. A single re-read of a 50MB transaction file across a slow corporate VPN tanks your close speed. Hot tier is for SPEED.

Tier 2 — Warm audit-ready (recent past, weekly access)

Lives on: object storage (S3, R2, or Backblaze B2) with versioning + write-once-read-many (WORM) lock for closed periods.

Contains: - Closed-period working files (the last 4-8 quarters) - Audit-ready close packages with embedded source data - Tax provision workpapers - LP report PDFs with the underlying calc data preserved

Retention: 7 years minimum — the typical LPA floor (the IRS general statute is 3 years; its worthless-securities rule reaches 7); some funds keep it longer.

Why this matters: an auditor in 2030 asking about a 2026 transaction needs to find the answer in MINUTES. WORM locks prevent accidental edits to closed periods. Versioning catches the case where someone edits a "closed" file they shouldn't.

Tier 3 — Cold archive (deep past, audit-defense only)

Lives on: deep glacier storage (S3 Glacier Deep Archive, or equivalent) with manual restore.

Contains: - Source data from prior years (raw exchange CSVs, API dumps, blockchain explorer exports) - Counterparty contracts in their original PDF form - Audit working papers from completed audits - Historical valuation files for ASC 820 fair-value memos

Retention: 10+ years.

Why this matters: in case of dispute, litigation, or extended audit, you need to reconstruct what was true at a point in time. Cold storage is INSURANCE — you hope you never need it but you absolutely need it to exist.

Where AI agent logs live

A category most fund admins forget. When AI agents run close work, they generate logs: tool calls, intermediate decisions, exceptions surfaced, reconciliation diffs. These logs are part of your audit trail.

Recommendation: route agent logs to Tier 2 (warm audit-ready) with the same retention as the working files they touched. Tag each log with the close period it served.

If you want your current storage mapped to the three-tier model — and the gaps identified — that's a 30-minute conversation.