Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.mnemom.ai/llms.txt

Use this file to discover all available pages before exploring further.

​
Provider support

Mnemom sits in front of three upstream model providers: Anthropic, OpenAI, and Gemini. Safe House, AIP integrity checkpoints, CLPI policy enforcement, and DLP all route through a single gateway, but the quality of each feature varies by provider — because the underlying APIs differ in what they expose. This page is the honest accounting: what Mnemom guarantees on each provider, what it partially guarantees, and what it cannot guarantee.
Where coverage is partial — particularly OpenAI’s integrity-checkpoint coverage — Mnemom does not claim parity. The v1 commitment is honest per-provider differentiation, not uniform coverage. Marketing materials, the status page, and the trust center all reflect this.

​
Supported models

Models the v1 promise applies to. Listed in the gateway’s /models.json registry; routed end-to-end through Safe House, AIP, CLPI, and DLP per the matrix below.

Anthropic

Claude Opus 4.7 Claude Sonnet 4.6 Claude Haiku 4.5

OpenAI

GPT-5 GPT-5 Codex o3 o3-mini

Gemini

Gemini 2.5 Pro Gemini 2.5 Flash
Canonical model IDs (what you pass in model: on a Mnemom API call): 
# Supported tier per ADR-051. Mnemom's gateway exposes this set via
# /models.json with `supported: true` flag; passthrough models route
# but carry no v1 promise.
supported_models:
  anthropic:
    - claude-opus-4-7
    - claude-sonnet-4-6
    - claude-haiku-4-5-20251001
  openai:
    - gpt-5
    - gpt-5-codex
    - o3
    - o3-mini
  gemini:
    - gemini-2.5-pro
    - gemini-2.5-flash
Additional models may route through the gateway in passthrough mode — they work for inference but do not carry the v1 feature-correctness promise. See Deprecation policy below.

​
Feature coverage matrix

Each cell describes the v1 commitment level. Symbols:
  • ✓ Fully supported. Tested in CI; Safe House features work identically to the Anthropic baseline.
  • ⚠ Partial — Provider exposes the feature, but with a documented limitation. See the supporting note.
  • N/A — Provider does not expose this capability. Not a Mnemom limitation.
CapabilityAnthropicOpenAIGemini
Streaming (SSE)✓✓✓
Tool use / function calling✓✓✓
Thinking-trace inspection (AIP)✓ [1]⚠ Partial [2]✓ [3]
Multimodal (image inputs)✓✓✓
Prompt caching✓ [4]⚠ Partial [5]⚠ Partial [6]
Batch API✓✓✓

​
[1] Anthropic — full extended thinking

Anthropic models expose full extended thinking blocks through the response API. AIP reads completed thinking blocks post-response, pre-delivery, and the verifier (Claude Haiku 4.5) has full chain-of-thought visibility. This is the most complete AIP coverage Mnemom offers — and the baseline against which other providers are compared.

​
[2] OpenAI — reasoning summaries only

OpenAI’s reasoning APIs (o3, o3-mini) expose reasoning summaries, not raw chain-of-thought. AIP can inspect these summaries, but the depth of analysis is bounded by what OpenAI chose to summarize. Non-reasoning models (gpt-5, gpt-5-codex, gpt-4o) expose no internal reasoning at all — AIP degrades to surface-only analysis on these. What this means in practice:
  • Boundary violations that surface in the model’s final response are caught equally well across all providers.
  • Boundary violations that would have surfaced in the model’s hidden reasoning, but never reach the final response, are caught with lower confidence on OpenAI than on Anthropic or Gemini.
  • The AIP integrity-checkpoint verdict is still emitted on OpenAI; the certainty of that verdict is provider-conditional.
This is the load-bearing honesty disclosure. Mnemom does not claim uniform AIP coverage across providers. See SLI-P1 in the Safe House SLOs for the public commitment.

​
[3] Gemini — full thoughts exposure

Gemini 2.5 models expose a thoughts field on response candidates. Coverage is uniform across 2.5 Pro and 2.5 Flash; AIP reads thoughts through the gateway’s response normalizer and treats it as equivalent to Anthropic extended thinking.

​
[4] Anthropic — explicit cache_control

Anthropic supports explicit cache_control block markers — customers control which prompt segments are cached. Mnemom passes cache_control through transparently. Safe House still evaluates the full request (it does not assume cached prefixes are safe just because they were previously seen). Cache hits do not bypass any checkpoint.

​
[5] OpenAI — automatic caching, no customer control

OpenAI’s prompt caching is automatic — the API decides what to cache based on request shape. Customers cannot reason about cache hit rates the way they can on Anthropic. Mnemom passes requests through unchanged; cache decisions are OpenAI’s. Safe House dispatch remains idempotent across cache hits and misses — the same prompt routed through Mnemom twice produces the same verdict regardless of whether OpenAI cached it.

​
[6] Gemini — separate CachedContent API

Gemini exposes prompt caching as a separate CachedContent resource (explicit cache lifecycle, named caches with TTL). The gateway today does not surface or use this API; requests are sent without referencing cached content. Customers using Gemini’s cache outside Mnemom will see lower latency than they see through Mnemom — this is a known gap, tracked in the Track 3 hardening plan.

​
AIP coverage by provider — the headline commitment

Of all per-provider gaps, integrity-checkpoint coverage is the load-bearing one. Mnemom’s public SLO commitment is:
ProviderTarget AIP coverage [*]
Anthropic (Opus 4.7, Sonnet 4.6, Haiku 4.5)≥ 99%
Gemini (2.5 Pro, 2.5 Flash)≥ 95%
OpenAI o-series (o3, o3-mini)≥ 50%
OpenAI non-thinking (gpt-5, gpt-5-codex, gpt-4o)Excluded — AIP runs in degraded “surface-only” mode
[*] “AIP coverage” is defined as the percentage of customer-facing AIP analysis cycles where the verifier had full thinking-trace inspection. Anthropic and Gemini expose full traces; OpenAI o-series exposes reasoning summaries only; non-reasoning OpenAI models expose no internal reasoning. See SLI-P1 for the rolling 30-day measurement methodology. If your application relies on AIP catching boundary violations in the model’s reasoning — particularly violations that would not surface in the final response — choose Anthropic or Gemini. OpenAI o-series is supported, with the documented coverage limitation. OpenAI non-thinking models are supported for inference but are not the right choice when AIP is the load-bearing safety layer.

​
Latency expectations

Per-provider P50/P95 latency for the Safe House dispatch path and AIP analysis is published on the status page and committed via SLI-P2 and SLI-P3. In aggregate:
  • Safe House dispatch adds ~15 ms P50 / ~60 ms P95 across all providers (the dispatch path is provider-agnostic).
  • AIP analysis runs post-response, pre-delivery. Cost varies by upstream-response token volume — Anthropic Opus with full extended thinking emits the largest traces and therefore the longest AIP analysis tails (P95 up to 2.5 seconds). Gemini 2.5 Flash and OpenAI non-thinking models emit thin traces and complete AIP analysis in P50 ≤ 800 ms.
Mid-stream interruption for integrity violations is not supported on any provider (per ADR-040). AIP runs after the upstream response completes, but before the gateway delivers it to the customer.

​
How we test against each provider

Mnemom’s gateway adapter — the code that parses each provider’s response format, extracts thinking blocks, and routes tool calls — is tested in two layers:
LayerCadenceCatches
Static-shapeEvery PR to the gateway”Did our parser break?” Asserts the adapter handles captured response fixtures correctly.
LiveNightly”Did the upstream provider ship a breaking change?” Real upstream call → real response → real parse.
All three providers (Anthropic, OpenAI, Gemini) run both layers as of 2026-05-12. Live tests fire nightly against real upstream APIs and detect breaking changes within 24 hours.
Streaming OpenAI thinking visibility: OpenAI’s STREAMING Chat Completions API does not emit reasoning summaries via delta.reasoning_content for o-series models. Reasoning happens server-side and is reported in usage.completion_tokens_details.reasoning_tokens, but the reasoning text itself does not reach the streamed response. Customers using OpenAI’s non-streaming Responses API see reasoning summaries; streaming customers do not.This is upstream behavior, not a Mnemom limitation. The AIP coverage commitment of ≥50% on OpenAI o-series applies to non-streaming consumers; streaming OpenAI o-series degrades to surface-only treatment. Mnemom’s gateway parser correctly reports empty thinking content for streaming OpenAI responses regardless of model.

​
Deprecation policy

Models the gateway routes are classified into two tiers:

Supported

Listed in the supported models section above. v1 promise applies. Safe House, AIP, CLPI, DLP all work to their per-provider commitment. Tested in CI. Deprecation requires 90 days’ notice via this page and the changelog.

Passthrough

Routed by the gateway but not in the supported tier. Inference works; Safe House features are best-effort. Not tested in CI. No deprecation notice — model availability tracks the provider’s lifecycle.
Today’s deprecation schedule:
ModelStatusSunset dateMigration target
claude-3-opus-20240229PassthroughTracks Anthropic’s deprecationclaude-opus-4-7
claude-3-5-sonnet-20241022PassthroughTracks Anthropic’s deprecationclaude-sonnet-4-6
claude-sonnet-4-20250514Passthrough2026-Q3 (recommend migrating now)claude-sonnet-4-6
gpt-4oPassthroughTracks OpenAI’s deprecationgpt-5
gemini-3-pro, gemini-3-flashPreviewWhen Google releases stableStay on gemini-3-* once supported
Passthrough-tier models that are removed from the upstream provider also disappear from the gateway’s /models.json registry; we do not maintain shims.

​
Out of scope

Provider expansion beyond Anthropic + OpenAI + Gemini is not in scope for v1. Cohere, Mistral, Together, Groq, and other providers are not supported. Adding a provider is a multi-quarter effort (Safe House dispatch, AIP adapter, CLPI policy schema mapping, harness coverage, docs) — tracked separately. BYOK (bring-your-own-key) for upstream providers is also out of v1. v1 ships with Mnemom-only key custody.