We Built an AI Agent That Runs 24/7, Never Forgets, and Checks Its Own Work

Memory + Scheduler + Decision Preflight. Three MCP servers that turn any agent from a session-bound chatbot into an autonomous worker. Here's exactly how it works — with real tool outputs from today.

The problem every AI agent has

AI agents are smart but forgetful. Every time a session ends — Claude Desktop restarts, Cursor closes, the browser tab dies — the agent loses everything. No user preferences, no project history, no compliance checks it ran yesterday, no portfolio decisions.

This isn't a minor inconvenience. It means agents can't do the things that actually matter for professionals: monitor something over time, remember what they learned, or work while you sleep.

We built three MCP servers that solve this. Together, they form what we call the autonomous agent stack: a brain, a motor, and a safety gate.

The autonomous agent stack

Each layer is a standalone MCP server. Together, they create a loop: the scheduler triggers, the gate checks, the tool executes, and memory stores the result. Next session, the agent has full context from day one.

Layer 1: MemoryOracle — persistent memory that survives restarts

MemoryOracle gives agents a namespace-isolated, searchable memory with SQLite FTS5 full-text search. It stores facts, preferences, decisions, action items, and session summaries — each categorized and importance-ranked.

Here's what it looks like when an agent stores a compliance fact:

{
  "status": "stored",
  "memory_id": "mem-c9fff3677d82",
  "namespace": "compliance-bot",
  "category": "fact",
  "importance": 8,
  "confidence": 1.0,
  "total_memories": 1,
  "timestamp": "2026-03-25T13:10:57Z",
  "oracle": "MemoryOracle"
}

Next session — hours, days, weeks later — the agent queries memory before acting:

FTS5 BM25 returns the stored fact instantly. No re-checking needed. The agent knows what it learned yesterday.

10 tools, 8 memory categories

Categories: fact preference decision observation rule session portfolio action_item

Layer 2: Decision Preflight — the compliance gate

Memory alone isn't enough. An agent that remembers everything but never checks whether it should act is dangerous. Decision Preflight runs five checks in parallel before any action: evidence verification, freshness, policy compliance, red-flag detection, and provenance.

One call. One verdict: go, caution, stop, or insufficient_evidence.

Watch what happens when an agent tries to recommend USDT to an EU client:

verdict:    STOP
confidence: 20%
blockers:
  - USDT is NOT AUTHORIZED under MiCA
  - USDT explicitly blocked by EU MiCA Compliance policy

The agent is prevented from making a wrong recommendation. MiCA enforcement is in 97 days. Without this gate, an agent trained on pre-2026 data might recommend USDT to EU clients — creating real legal risk.

Built-in policies

Layer 3: SchedulerOracle — the motor that runs while you sleep

The brain remembers. The gate checks. But who presses the button? SchedulerOracle is the motor — it triggers actions on a schedule, even when no human is present.

Schedule a daily MiCA compliance check:

task_id:  task-f159b5752d
schedule: cron 0 9 * * * (daily at 9:00 UTC)
target:   preflight/decision_preflight
next_run: 2026-03-26T09:00:00Z
memory:   auto-store enabled (importance: 8)

Tomorrow at 9:00 UTC, the scheduler will automatically run Decision Preflight on RLUSD, store the result in MemoryOracle, and be ready for the next session. The agent picks up exactly where it left off.

The complete loop — how it works in practice

Here's a real scenario we tested today:

Day 1, 10:00 AM — Agent runs MiCA compliance checks on 5 stablecoins. Stores results in memory. Schedules daily monitoring for RLUSD (pending authorization). Blocks USDT in the policy engine.

Day 2, 9:00 AM — Nobody is online. The scheduler fires. Preflight checks RLUSD: still CAUTION (pending). Result is stored in memory automatically.

Day 2, 2:00 PM — User opens a new session. Agent queries memory: "What do I know about RLUSD?" — instantly gets the 9am check result. No re-running. No waiting. No re-explaining.

Day 30 — RLUSD gets MiCA authorization (hypothetical). The scheduler catches it the same morning. Memory stores the change. Next session, the agent says: "RLUSD was authorized on March 28. Portfolio allocation can now include RLUSD at up to 10%."

What makes this different from other memory MCPs

Several memory MCP servers exist — mcp-memory-service, Hindsight, Engram, MemoClaw, OpenMemory. They solve the storage problem. But none of them have:

The difference isn't memory alone — it's memory + scheduler + compliance gate as one integrated stack. That's what turns a chatbot into an autonomous worker.

x402 pricing — pay per decision, not per month

Every tool is available via x402 pay-per-call using USDC on Base. No account, no API key. The agent pays per action.

A daily compliance monitoring agent costs roughly $0.20/day — one preflight + one memory store + one query. That's $6/month for 24/7 autonomous compliance monitoring.

Connect in 30 seconds

All three servers are live. Add them to Claude Desktop, Cursor, Windsurf, or any MCP client:

Or call directly via HTTP:

Who is this for

Compliance teams — Schedule daily MiCA/DORA checks, store results, get notified on changes.
Portfolio managers — Track stablecoin allocations over time, remember risk decisions.
Research agents — Build knowledge bases that persist across sessions.
Procurement — Remember vendor assessments, schedule re-evaluations.
Any agent that needs to work longer than one conversation.

What we learned building this

Memory without scheduling is a notebook. It's useful, but it doesn't act. You still need a human to open it.

Scheduling without a gate is dangerous. An autonomous agent that runs compliance checks but can't distinguish between PASS and FAIL will eventually make a costly mistake.

All three together is what makes it work. The scheduler is the motor. Memory is the brain. Preflight is the safety check. Remove any one and the system degrades from "autonomous worker" to "fancy tool."

Frequently Asked Questions

Source code: MemoryOracle · SchedulerOracle · Decision Preflight · Trust Layer
Platform: tooloracle.io · x402: tooloracle.io/x402/ · Status: Memory · Scheduler · Preflight