Mnemom Gateway Overview

​

Architecture

gateway/
├── cli/          # CLI tool (npm package)
├── gateway/      # API proxy + tracing
├── observer/     # Trace builder + AAP verification
├── api/          # REST API
└── database/     # Postgres schema (self-hosted: Supabase)

​

Request flow

                    ┌─── /anthropic/* ──→ Anthropic (Claude)
Your App → mnemom   ├─── /openai/*    ──→ OpenAI (GPT-5)
           gateway  └─── /gemini/*    ──→ Google (Gemini)
                ↓               ↑
           request logging ← SSE stream
                ↓           |
         Response split     |
           /       \        |
    to client    to background analysis
   (immediate)   (async)
                    |
              Parse SSE → extract thinking
                    |
              AIP analysis (Haiku, ~3s)
                    |
           checkpoint + attestation stored
                    |
           Policy evaluation (autonomy_mode-gated)
                    |
           Observer (cron) → AP-Trace stored
                    |
           Dashboard (mnemom.ai)

1. Gateway — Intercepts API requests to Anthropic, OpenAI, and Gemini. It identifies your agent via API key hash (zero-config), attaches tracing metadata, and injects thinking/reasoning per provider. What the gateway does with the analysis depends on the integrity mode. In observe and nudge modes, streaming responses (99%+ of traffic) are returned to the client immediately (zero added latency) while AIP analysis runs in the background — extracting thinking blocks, running integrity analysis with Claude Haiku (~3s), creating cryptographically attested checkpoints (Ed25519 + hash chain + Merkle tree), and — in nudge mode — recording the verdict for carryover into the agent’s next turn. Non-streaming responses in these modes are analyzed post-hoc after the response returns. In enforce mode the guarantee is stronger: the gateway gates the same turn on both streaming and non-streaming, evaluating the response before it is delivered so a violating response is blocked rather than delivered — on streaming this adds latency. That describes the response path. On the request path the gateway also runs Safe House front-door screening, and that is not a single gate per turn: it screens the inbound message and, when screen_surfaces.tool_responses is enabled, each tool result the request is handing back to the model — synchronously, before the body is forwarded upstream. A flagged tool result is withheld (replaced with a quarantine notice) or decorated as untrusted in that same request. See When the front door runs.
2. Observer — A scheduled background process that processes gateway logs to build AAP-compliant AP-Traces. It extracts thinking blocks from logged responses, analyzes decisions with Claude Haiku, builds AP-Traces, and verifies them against alignment cards using the AAP SDK. The observer links its traces to gateway-created AIP checkpoints. AIP integrity analysis is handled entirely by the gateway.
3. API — Serves agent data, traces, integrity scores, drift alerts, enforcement status, and a unified conscience timeline. Powers both the CLI and the web dashboard.
4. CLI — The mnemom command. Configures your local environment and queries your agent’s transparency data. See the CLI Reference for details.
5. Dashboard — Web UI at mnemom.ai where you can view the conscience timeline, claim your agent, and monitor alignment.

​

Supported providers

​

What gets traced

• Action — What the agent did (type, name, category)
• Decision — What alternatives were considered and why one was selected
• Escalation — Whether the agent escalated to a human and why
• Verification — Whether the trace is consistent with the agent’s declared alignment card
• Integrity — Per-turn AIP analysis of thinking blocks, with verdict (clear/review_needed/boundary_violation)
• Protection — Safe House front-door and back-door verdicts. The front door records one evaluation per inbound surface, so a turn that carries tool results produces several — see When the front door runs
• Policy — Policy evaluation results including verdict (pass/warn/fail), violations, warnings, and coverage report. The X-Policy-Verdict response header indicates the result.

​

What is NOT stored

​

AIP compatibility

​

Claiming your agent

1. Run mnemom status to get your agent ID
2. Go to mnemom.ai/claim
3. Paste your agent ID and prove ownership with your API key hash

​

Current limitations

• API key auth only — Agent identification uses API key hashing (SHA-256, works identically across all providers). OAuth and other auth methods are not supported.
• Multi-agent support — Multiple named agents can share a single API key with distinct identities using the x-mnemom-agent header. See the Multi-Agent Setup guide and the Agent CRUD API for programmatic registration.
• AIP requires exposed reasoning — AIP integrity checking needs a provider that exposes thinking content. Anthropic and Gemini do; OpenAI’s Chat Completions door does not expose any reasoning content today, on any model, so OpenAI gets synthetic clear verdicts rather than analysis — see Provider Support.
• Hosted gateway — The default gateway runs on Mnemom infrastructure. Self-hosting is possible but requires manual setup.

​

Dependencies

• Agent Alignment Protocol (AAP) — @mnemom/agent-alignment-protocol on npm
• Agent Integrity Protocol (AIP) — @mnemom/agent-integrity-protocol on npm
• API keys: Anthropic (required for AIP analysis), OpenAI and Gemini (optional, for multi-provider tracing)

​

Next steps