@onemem/openai-agents

Record OpenAI Agents SDK runs as verifiable on-chain OneMem TraceSessions (Sui + Walrus + Seal) — every tool call + the final output captured as Merkle-chained ActionCalls anyone can verify.

Publication note, 2026-06-18: @onemem/[email protected] is current on npm after pnpm registry:status --strict and includes createOneMemMemory(...). Re-run that command before making a fresh public install claim.

Usage

import { Agent } from "@openai/agents";
import { createOneMemMemory, createTracedRunner } from "@onemem/openai-agents";

const agent = new Agent({ name: "research", instructions: "..." });

// A Runner with OneMem trace recording attached.
const runner = createTracedRunner({ agentId: "my-app" });

// Optional explicit memory helper — recall before a run, capture after.
const memory = createOneMemMemory();
const input = await memory.recallInto("Find the latest on X and summarize.");
const result = await runner.run(agent, input);
await memory.capture(`User asked: ${input}\nAgent answered: ${result.finalOutput}`);
// → each tool call + the final answer recorded as one verifiable trace.

Already have a Runner? Attach to it:

import { Runner } from "@openai/agents";
import { attachOneMem } from "@onemem/openai-agents";

const runner = attachOneMem(new Runner(), { agentId: "my-app" });

How it works

Listens to the Runner lifecycle events (agent_tool_start/agent_tool_end/ agent_end), buffers each run's calls per RunContext (so concurrent runs are isolated), and flushes one TraceSession via @onemem/sdk-ts/runtime's recordSession at agent_end. Recording is fire-and-forget — it never slows or breaks the run; a OneMem failure is logged and swallowed.

Zero-config: namespace/cap/signer auto-provision on first use (persisted under ~/.onemem/). Override via options or env (ONEMEM_NAMESPACE_ID + ONEMEM_RW_CAP_ID, ONEMEM_PRIVATE_KEY, SUI_NETWORK).

Options

agentId / environment (default "openai-agents"), network, privateKey, target, label, enableTrace (default true), onTrace(sessionId), logger.

Memory helper

createOneMemMemory() exposes the Mem0-style side of OneMem for OpenAI Agents apps:

• recall(query, topK?) searches configured MemWal memories.
• recallInto(input, topK?) prepends recalled memories as a context block and returns input unchanged when memory is disabled or no memories match.
• capture(text) stores a durable memory through MemWal and returns true on success, false when disabled or failed.

Memory is explicit because it changes the agent input. Runner event listeners remain responsible for trace capture only. Configure memory with the same credentials used by the SDK memory layer: delegate key, MemWalAccount id, embedding API key, MemWal package id, and relayer URL. Local onemem login credentials and env vars are both supported by the shared runtime resolver.

Scope (v0.1)

Trace capture is automatic through Runner lifecycle events. In repo-local source, memory recall/capture is shipped as the explicit createOneMemMemory() helper. Verify/replay tools and automatic memory extraction remain tracked follow-ups. Requires @openai/agents v0.11+ (peer dependency). Spec: docs/05-our-architecture/04-frameworks/openai-agents-tools.md.