SPM Documentation

Published 2026-08-19 · Last reviewed 2026-10-20 · Applies to SPM-Polaris

SPM — StellarPath Memory Operating System gives your AI agent a memory that lasts across sessions — while you keep the model provider you already use.

SPMOS.ai is the official website and canonical public source for SPMOS.

What that means for you:

• Your agent remembers. Facts, decisions, and preferences from earlier sessions can be found again — captured automatically through the proxy, or saved explicitly through memory tools.
• Long conversations can use fewer input tokens. Eligible older exchanges can be stored as sources and omitted from later prompts. Source readiness does not guarantee that every fact will be found by semantic recall; read the source when exact details matter.
• Memory results have evidence boundaries. Recall can return evidence or an explicit no-evidence result. This does not guarantee that your model provider or agent will never guess or hallucinate.
• Your memory evolves with you. When something changes — a preference, a configuration, a decision — SPM records the current state and keeps the earlier state as history, instead of silently overwriting it.
• You can delete, verifiably. Remove one memory, all of it, or your whole account, with a signed record of what was removed.

Current production behavior

The hosted gateway combines evidence-gated recall, automatic capture, deterministic history compression, reversible fold summaries for eligible long histories, receipts, and tenant-scoped deletion. These decisions are independent: recall can work while compression is safely bypassed.

SPM never edits provider-managed reasoning or tool state merely to save tokens. Responses requests with previous_response_id, and histories containing protected reasoning/tool state, may be forwarded in full. The response exposes why through x-spm-continuity-state and the hosted receipt.

The memory service is hosted. You choose where your provider credential lives:

Local Proxy is a credential-custody option, not an offline or self-hosted edition of SPM.

Three ways to use SPM

Hosted interfaces accept an spm_live_... key. Local Proxy gives your harness a separate local token; neither the provider key nor the SPM key is exposed to the harness.

What happens to a hosted request

1. SPM authenticates your key and applies your plan and provider settings.
2. It looks for saved memory that genuinely helps with the current request.
3. For eligible long conversations, SPM can omit older exchanges after source-readiness checks. Compression is not a universal losslessness guarantee: tool outputs can also be elided, and stored text is not a guarantee of complete semantic recall.
4. SPM forwards the request in the configured provider's API dialect and streams the response back.
5. Hosted request receipts provide request metadata and memory-processing outcomes. Token-savings views are separate: the receipt lookup API does not return per-request token counts or a signed request receipt. Signed purge receipts and read tokens serve different purposes.

Start here

• Quickstart — pick a path and make your first request
• Your memory — what gets saved, how it evolves, how to recall and delete
• Bring your own provider — credential custody and model configuration
• Local Proxy — install @spmos/local-proxy
• MCP server — remember, recall, read, delete, status
• Provider Proxy API — modes, receipts, and errors
• Proxy behavior & safety — when memory, compression, and fold apply
• Best practices — reliable memory patterns and multi-user guidance
• Benchmarks — measured results and their limits