SPM Documentation
SPM (StellarPath Memory) is memory and context infrastructure for AI agents, delivered as a hosted provider proxy with a governed long-term memory plane. You point any OpenAI- or Anthropic-compatible client at the SPM endpoint and keep your existing LLM provider account; every request then gains evidence-gated recall, deterministic context compression, and a signed audit trail — with no generative LLM call added to the memory path.
SPM is a hosted service. There is no appliance or self-deployment: you integrate through the provider proxy or the MCP server while we operate the memory plane.
Two doors in
| Door | Endpoint | For |
|---|---|---|
| Provider proxy | https://api.spmos.ai/v1 |
Any OpenAI/Anthropic-compatible app or agent |
| MCP server | https://api.spmos.ai/mcp |
Codex, Claude Code, and other MCP clients |
Both doors accept the same spm_live_... API key and enforce the same tenancy, quota, and evidence-gate rules.
What SPM does inside a request
- Authenticates your API key and resolves your tenant, namespace, policy, and provider channel.
- Recalls relevant memory through lexical and vector legs, fused and admitted through an evidence gate. If nothing qualifies, the gate fails closed instead of guessing.
- Compresses the assembled context deterministically. The memory path does not call a generative LLM, so memory growth does not increase inference cost.
- Forwards to your provider in its native dialect (OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages) and streams the response back.
- Ingests the exchange asynchronously through idempotent, retried, dead-lettered jobs, then writes a signed request receipt with full token accounting.
Recall and forwarding are explicitly separate stages: your provider never decides what is remembered, ranked, admitted, or compressed.
Plans at a glance
| Free | Starter | Growth | Enterprise | |
|---|---|---|---|---|
| Price | $0 | $9/mo · $99/yr | $15/mo · $150/yr | Custom |
| API rate limit | 10 req/min | 30 req/min | 100 req/min | Reviewed |
| Memory writes / month | 500 | 3,000 | 8,000 | Reviewed |
| Stored memories | 1,000 | 10,000 | 50,000 | Reviewed |
| Provider channels | 1 | 3 | 10 | Reviewed |
Full terms and enforcement behavior: Plans & quotas.
Where to go next
- Quickstart — first memory-backed call in five minutes
- Authentication & API keys — keys, scopes, and hygiene
- Plans & quotas — pricing, allowances, and limit behavior
- Bring your own provider — presets, live model catalog, custom endpoints
- Provider proxy API — endpoints, receipts, error codes
- Memory, evidence & deletion — the gate, provenance, and signed purge receipts
- MCP server — agent-native memory tools
- Architecture — planes, storage doctrine, failure behavior
- Benchmarks — the scorecard, with footnotes