- Per-operation (pay-per-call): e.g., $0.001 per store, $0.001 per recall — simple, scales naturally with usage, easy for developers to reason about cost, but can feel expensive at high volume without volume discounts.
- Subscription tiers: fixed monthly price for a usage band (e.g., up to N memory operations) — predictable revenue for you, predictable cost for customers, but requires you to actually model your own infrastructure cost per operation accurately to avoid pricing yourself into a loss on heavy users.
- Token/usage-based: pricing tied to underlying LLM token consumption (since extraction and routing calls are your actual variable cost) — most accurately reflects your true costs, but is harder for customers to predict/budget against.
The friction in your signup → first successful API call flow is often a bigger factor in developer adoption than your actual retrieval quality, especially early on when developers are just evaluating whether to try you at all. The bar being set in this space is instant, frictionless provisioning — no email verification loop, no dashboard clicking, a working API key in seconds. Match or beat this bar, or developers will bounce before ever testing your actual differentiation.
Prioritize, in order: (1) a raw REST API (works for literally anyone, any language), (2) an MCP server (works with Claude Code, Cursor, and the growing MCP ecosystem with zero custom integration), (3) SDKs for the 2-3 most popular agent frameworks in your specific use case (not all frameworks — the ones your actual target customers use).
Your pitch should never be "we also do memory" — that's a losing frame against an established generalist. Instead: "we solve [specific, named failure mode] for [specific use case] agents, because we understand [specific domain fact type / decay pattern / contradiction rule] that generic memory tools treat as one-size-fits-all." Back this with your Module 12 evaluation harness — a measured accuracy number on your agent use case-specific benchmark is a far stronger technical argument than a generic claim of "better memory."
- Write your system summary in exactly 2-3 sentences, naming the specific failure mode you solve and the specific use case.
- Sketch a pricing table with at least 2 tiers, and calculate your actual per-operation cost (LLM API cost for extraction + routing + your infra cost) to sanity-check that your proposed pricing has real margin.
- List, in priority order, the first 3 integrations (REST API always first) you'd build, and justify the order based on where your actual target users already work.