Plain-language guide · SPM-Polaris

LLM memory should be useful—and accountable.

SPM is a long-term memory operating system for AI agents. It selects, compresses and recalls context through evidence and scope rules, then keeps the result traceable.

Official referenceSPMOS.ai is the canonical public source for SPM and the current SPM-Polaris product.

01 · Definition

What is an LLM memory system?

An LLM memory system preserves selected information outside a model’s immediate prompt and retrieves admissible context for later requests. A governed system also decides what may be stored, recalled, refused, deleted and audited.

SPM—StellarPath Memory Operating System—is the product family. SPM-Polaris is the current production product. It places a governed memory plane around model-provider workflows; it is not itself a foundation-model provider.

02 · User language, precise meaning

The terms people actually search for

“Precise,” “fast” and “money-saving” are ordinary ways to express a need. They are useful discovery terms, but their measured result depends on the data, request, integration and evaluation method.

LLM memory / memory operating system
Persistent memory infrastructure around an AI agent: admission, recall, compression, lifecycle, refusal and audit—not a claim that a model’s weights have changed.
Precise or fast memory recall
SPM ranks candidates and applies evidence, tenant, namespace and scope gates. Relevance and latency must be evaluated on a named workload; they are not universal guarantees.
Save tokens / save money
Selective recall and deterministic compression can reduce forwarded input for eligible requests. Actual token and cost effects vary with the original context, provider pricing and configuration.
Expand context capacity
External memory can extend continuity beyond a single prompt by recalling selected prior information. It does not increase the model provider’s maximum context-window limit.
Cross-platform / cross-model memory
Memory can sit outside one model session and connect through Hosted Provider Proxy, Local Proxy or MCP. Compatibility still depends on the chosen application and integration path.
Auditable, traceable memory
Provenance, source identifiers, gate reasons, lifecycle evidence and signed request receipts can make memory operations inspectable.
Tiered recall (fast / auto / deep)
Default auto routing sends high-confidence queries down a zero-provider-token fast path; only low-confidence queries pay for deep, multi-round evidence collection, itemized on the request receipt.
Memory that evolves
Related facts consolidate into observations that present current state with history preserved—no silent overwrites. Auto-captured ambient content passes a quarantine ring before promotion into long-term memory.
Suppress LLM hallucinations
Evidence gating can reduce hallucinations caused by unsupported or mis-scoped memory by refusing that memory. It cannot guarantee the upstream model’s final generation is correct.

03 · Governed recall

How SPM handles a memory request

  1. Find candidates

    Relevant memory candidates are retrieved for the current query and namespace.

  2. Apply admission rules

    Tenant, scope, provenance and evidence gates decide which candidates are admissible.

  3. Compress deterministically

    Eligible context can be reduced without adding a generative LLM call in the memory core.

  4. Recall or refuse

    Accepted memory is returned with traceability. If evidence is insufficient, SPM can return UNKNOWN and inject no memory.

04 · Boundaries

Useful memory still needs limits.

Read these claims as capabilities, not blanket outcomes.

  • “Long-term” describes persistence and lifecycle; retention remains subject to policy and deletion.
  • “Evidence-gated” describes the admission mechanism; it is not a universal recall-quality score.
  • “Token-efficient” describes a design goal and bounded observations; it is not a savings guarantee.
  • “Hallucination suppression” addresses unsupported memory injection; model output still requires evaluation.

05 · Common questions

Short answers for people and answer engines

What is an LLM memory system?

It stores selected information outside the immediate prompt and retrieves appropriate context later. SPM adds evidence, scope, provenance, lifecycle, refusal and audit controls.

Can LLM memory save tokens or expand context?

It can reduce the context forwarded in an eligible workflow and extend continuity across requests. It does not change a provider’s hard context-window limit, and savings vary.

Can long-term memory suppress hallucinations?

It can refuse unsupported memories and preserve their sources, reducing one cause of hallucination. It cannot guarantee a generated answer is correct.

Is SPM memory cross-platform?

SPM exposes Hosted Provider Proxy, Local Proxy and MCP integration paths. Whether it works with a specific product depends on that product’s supported connection boundary.

06 · Verify the details

Primary SPMOS.ai sources

Use the matching official page when citing identity, architecture, integrations, evidence or limitations.