REST / OpenAPI
Use the public, versioned HTTP contract from any trusted backend stack.
HTTPS · JSON · api.tmcra.comOpen integration guide Developer integrations / 01
Connect TMCRA to the agent stack you already run.
Use the stable HTTP contract directly, wrap your model call with a preview SDK lifecycle, or connect a native host adapter. The labels below describe release availability, not merely whether source code exists.
02 / INTEGRATION SURFACES
Start at the layer that matches your system.
Stable means the public contract is supported. Preview means the integration is testable but still evolving. Pilot means access is reviewed and the native adapter is not yet a generally published package.
Python SDK
Typed sync and async clients with an optional recall → answer → write lifecycle wrapper for your own Agent runtime.
SyncMemoryLifecycle · AsyncMemoryLifecycleOpen integration guide JavaScript / TypeScript SDK
One compiled package for JavaScript and TypeScript, with an optional automatic turn wrapper and multi-Agent attribution.
TMCRAMemoryLifecycle · @tmcra/typescriptOpen integration guide MCP Server
Explicit recall and ingest tools over local stdio. Automatic turns require lifecycle support from the host, such as the Codex Hooks mode.
explicit MCP · optional Codex HooksOpen integration guide Codex
A downloadable preview that recalls before Codex answers, records completed main and subagent turns, isolates projects, and can import retained Codex history.
SessionStart · SubagentStart · UserPromptSubmit · Stop · SubagentStopOpen integration guide DeepSeek Harness
Install the Harness package, sign in to TMCRA through device authorization, then recall before each answer and write completed user and Agent turns back with project and multi-Agent provenance.
account login · automatic recall · role-separated writebackOpen integration guide OpenClaw
Native Hooks automatically recall on the current prompt and capture the completed turn. Specialized Agents share project memory without losing Agent or session attribution.
before_prompt_build · agent_endOpen integration guide Hermes Agent
The native MemoryProvider recalls before an answer, queues completed turns, and records delegated Agent work without mislabeling it as user speech.
prefetch · sync_turn · on_delegationOpen integration guide 03 / PRODUCTION FLOW
The question comes before recall; the answer comes before write.
- 01
Keep credentials out of prompts
Store the API Key in a protected local or server-side environment. Never place it in a prompt, browser bundle or project source file.
- 02
Define stable memory scopes
Separate user-global memory from project memory, then retain each conversation as a session inside its project.
- 03
Recall from the current question
After the user submits a prompt and before the model answers, recall allowed scopes and inject bounded prompt_evidence.content.
- 04
Write the completed turn idempotently
Persist user and assistant messages separately in shared project memory, preserving role, Agent, session, and source provenance.
04 / AUTOMATIC LIFECYCLE
One project can contain many Agents.
Use a shared project Scope across specialized Agents, distinct Sessions for their conversations, and explicit actor provenance on every record. The optional Agent-private layer is recall-only and off by default.
Read the automatic memory and multi-Agent guideturn.lifecycle
user.prompt
→ recall + inject
→ agent.answer
→ write(user, assistant)05 / DELIVERY CONTRACT
Your model stays yours.
TMCRA does not replace the model or Agent runtime. It supplies traceable evidence and a deterministic prompt context at the point where your system needs memory.
recall.response
{
"evidence": [...],
"prompt_evidence": {
"content": "..."
}
}06 / PILOT
Tell us what you need to connect.
Submit your stack, use case and rollout timeline. The TMCRA team will review the application in the Internal control plane.