OpenAI-compatible gateway

UnifiedMemory exposes OpenAI-style routes for clients that already know how to call chat completions or responses APIs.

POST /v1/chat/completions
POST /v1/responses

Use these when you want familiar model-client shapes with UnifiedMemory auth, scope, and optional memory context.

Auth

Authorization: Bearer $UNIFIED_MEMORY_API_KEY

Gateway keys should be scoped. The key decides container, source app, agent identity, and available tools.

Chat completions example

Common request fields:

Field	Required	Type	Description
model	Yes	string	Provider/model route. MiniMax routes should use configured model names.
messages	Yes for chat completions	array	OpenAI-style chat messages.
input	Yes for responses	string or array	OpenAI Responses-style input.
memory	No	object	UnifiedMemory options such as container_tag, trust_policy, and max_context_tokens.
stream	No	boolean	Streaming hint when supported by the gateway route.
metadata	No	object	Caller metadata. Do not include secrets.

curl https://edge-api.jithendranara.dev/v1/chat/completions \
  -H "Authorization: Bearer $UNIFIED_MEMORY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/MiniMax-M2.7",
    "messages": [
      {"role": "system", "content": "Use memory only when supplied."},
      {"role": "user", "content": "What should I know before editing the docs?"}
    ],
    "memory": {
      "container_tag": "jeethendra",
      "trust_policy": "balanced",
      "max_context_tokens": 900
    }
  }'

Responses example

curl https://edge-api.jithendranara.dev/v1/responses \
  -H "Authorization: Bearer $UNIFIED_MEMORY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/MiniMax-M2.7",
    "input": "Summarize the current memory docs rules.",
    "memory": {
      "container_tag": "jeethendra",
      "include_courtroom_links": true
    }
  }'

Response shape

The gateway preserves the client-compatible response envelope and may add memory diagnostics under metadata or response extensions:

{
  "id": "resp_01",
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Before editing docs, keep public pages curated and exclude proof artifacts."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 620,
    "completion_tokens": 28,
    "total_tokens": 648
  },
  "memory": {
    "container_tag": "jeethendra",
    "tokens_injected": 210,
    "trust_policy": "balanced",
    "cross_agent_candidates_suppressed": 0
  }
}

Common outcomes

Status	Meaning
200	Model response returned.
401	Missing or invalid key.
403	Key lacks gateway or memory capability.
404	Model, container, or route unavailable.
422	Invalid OpenAI-style request payload.
503	Upstream model or cognition provider unavailable.

Boundaries

• Gateway memory context still follows trust and scope gates.
• Test/eval/forensic rows are not injected because a chat client asked for them.
• Use Agent Context when you want the memory packet without model synthesis.