reloop-ai 0.3.2

Failure memory for AI agents — self-healing retry with structured learning

ReLoop

Failure memory for AI agents.

Every agent fails. ReLoop is the first framework that gets smarter from failure.

---

The Problem

AI agents retry blindly -- same mistake, same failure, burning tokens and money. No framework treats failure as data. They either retry with no memory, or give up.

The Solution

ReLoop captures every failure into a structured memory graph -- error type, root cause, suggested fix, confidence score, semantic embedding -- so the next retry starts smarter. Your agents don't just recover. They get permanently smarter.

---

Quick Start

The reloop-ai package is officially published and ready for production use.

pip install reloop-ai
reloop init
reloop demo

Minimal Setup (No Redis Required)

ReLoop works out of the box with SQLite -- no Redis or Blaxel needed:

pip install reloop-ai
export OPENAI_API_KEY=sk-...
reloop serve

Redis and Blaxel are optional -- add them when you need production performance. For local development, ReLoop uses Docker or runs commands directly as the sandbox.

---

Three Ways to Use ReLoop

1. As a Library (any agent, 3 lines)

from reloop import FailureMemory

memory = FailureMemory(redis_url="redis://localhost:6379")
similar = await memory.search("ImportError sharp")  # Returns past failures + fixes

2. As a Framework (full self-healing loop)

reloop run "Fix and deploy the Next.js project at ./my-broken-app"

3. As an MCP Server (Cursor & Claude Code)

Integrate ReLoop directly into your AI IDEs so they become self-healing. When the MCP server is connected, Claude can:

• Search Memory: Semantically search past errors across your team's history.
• Fetch Checkpoints: Restore exact state from past Blaxel Firecracker VM checkpoints.
• Execute Safely: Run isolated code tests within Blaxel sandboxes directly from the editor.

Add to your claude_desktop_config.json or Cursor settings:

{
  "mcpServers": {
    "reloop": {
      "command": "python",
      "args": ["-m", "reloop.mcp_server"]
    }
  }
}

---

Architecture

┌─────────────────────┐   MCP Protocol    ┌──────────────────────┐
│  Cursor / Claude    │ ───────────────►  │  ReLoop MCP Server   │
└─────────────────────┘                   └──────────┬───────────┘
                                                     │
                               ┌─────────────────────┴──────────────────────┐
                               │                                             │
                               ▼  Vector Search                              ▼  Isolated Execution
                    ┌──────────────────────┐                    ┌──────────────────────┐
                    │  Redis Agent Memory  │                    │  Blaxel Firecracker  │
                    └──────────┬───────────┘                    └──────────┬───────────┘
                               │ stores                                    │ 25ms resume
                               ▼                                           ▼
                    ┌──────────────────────┐                    ┌──────────────────────┐
                    │  Failure Embeddings  │                    │   State Checkpoints  │
                    └──────────────────────┘                    └──────────────────────┘

Redis: The Memory Backbone

ReLoop utilizes Redis as the core of its Agent Memory Server. The architecture consists of a 3-tier memory system:

1. Working Memory: Stores the current task's session state and immediate context.
2. Long-term Memory: A persistent failure graph using Redis Vector Search to semantically match current errors with past distilled solutions.
3. Episodic Memory: Full execution traces and timeline records for auditing and the dashboard UI.

Blaxel: Perpetual Execution Sandboxes

For safe, deterministic task execution, ReLoop integrates Blaxel's Firecracker microVMs.

• Perpetual State: The sandbox is never lost. You can pause and resume the exact environment.
• 25ms Checkpoint/Restore: ReLoop creates instantaneous checkpoints after every step.
• Time-Travel Rewinds: Hit a roadblock? Rewind the agent to a previous checkpoint in 25ms and try a different fix strategy.

The REJD Loop

The core algorithm: Retrieve -> Execute -> Judge -> Distill

                           ┌─────────────┐
                           │  New Task   │
                           └──────┬──────┘
                                  │
                                  ▼
                    ┌─────────────────────────────┐
                    │           Retrieve           │
                    │  Query Redis for similar     │
                    │      past failures           │
                    └──────────────┬──────────────┘
                                   │
                                   ▼
                    ┌─────────────────────────────┐
                    │           Execute            │
                    │    Run in Blaxel sandbox     │
                    └──────────────┬──────────────┘
                                   │
                                   ▼
                    ┌─────────────────────────────┐
                    │            Judge             │
                    │     Success or failure?      │
                    └──────┬──────────────┬────────┘
                           │              │
                        Success        Failure
                           │              │
                           ▼              ▼
              ┌────────────────┐  ┌───────────────────────┐
              │ Distill        │  │ Distill Failure        │
              │ Success        │  │ root cause, fix,       │
              │ Store solution │  │ confidence score       │
              └───────┬────────┘  └──────────┬────────────┘
                      │                       │
                      ▼                       ▼
              ┌────────────────┐  ┌───────────────────────┐
              │ Task Complete  │  │  Circuit breaker or    │
              └────────────────┘  │   budget exceeded?    │
                                  └──────┬──────────┬──────┘
                                         │          │
                                        Yes         No
                                         │          │
                                         ▼          └──► Retrieve ↑
                                ┌─────────────────┐
                                │ Task Abandoned  │
                                └─────────────────┘

Powered by:

• OpenAI Agents SDK -- orchestrates the REJD loop with handoffs between specialist agents
• Redis -- 3-tier failure memory (working, long-term, episodic) via Agent Memory Server
• Blaxel -- Firecracker sandbox with 25ms checkpoint/restore (optional; Docker or direct execution for local dev)

---

Timeline UI

The non-chat interface that makes failure learning visible.

A horizontal timeline of colored nodes tells the full story at a glance:

RED (failed) -> RED (failed) -> RED (failed) -> GREEN (succeeded)

Click any node to inspect the full failure record -- root cause, suggested fix, confidence score, cost, and the exact code diff that resolved it.

---

Integrations

Works with any agent framework:

• OpenAI Agents SDK
• LangGraph
• CrewAI
• Claude Agent SDK
• Raw Python

ReLoop is the memory layer -- bring your own orchestration.

---

A/B: Memory vs No Memory

Metric	Without Memory	With Memory
Attempts to fix 4 bugs	12+	4
Total cost	$0.47	$0.18
Same mistake repeated	3x	0x

---

API Reference

Full API reference: docs/api-reference.md

Method	Path	Description
POST	/v1/tasks	Create and run a task
GET	/v1/tasks/{id}	Get task status and result
GET	/v1/tasks/{id}/timeline	Full execution timeline
GET	/v1/tasks/{id}/sse	Server-Sent Events stream
POST	/v1/memories/search	Semantic search over failure memory
GET	/v1/memories/stats	Aggregated memory statistics
GET	/v1/tasks/{id}/checkpoints	List sandbox checkpoints
POST	/v1/tasks/{id}/checkpoints/{cid}/restore	Rewind to checkpoint
POST	/v1/tasks/ab-comparison	Run A/B comparison (with vs without memory)
GET	/v1/tasks/{id}/circuit-breaker	Get circuit breaker state for a task
POST	/v1/memories/predict	Predict failure likelihood for new code
GET	/v1/memories/export	Export failure memory as JSON

---

Configuration

ReLoop is configured via environment variables. Only OPENAI_API_KEY is required -- everything else has sensible defaults.

Variable	Required	Default	Description
OPENAI_API_KEY	Yes	--	OpenAI API key for code generation and reasoning
REDIS_URL	No	redis://localhost:6379	Redis connection URL (falls back to SQLite)
REDIS_MEMORY_INDEX	No	reloop-failures	Vector index name for failure embeddings
BLAXEL_API_KEY	No	--	Blaxel API key for Firecracker sandboxes
BLAXEL_WORKSPACE	No	--	Blaxel workspace name
CODEX_MODEL	No	gpt-4o	Chat model for planner/distiller
REASONING_MODEL	No	o1	Deep reasoning model for root cause analysis
FAST_MODEL	No	gpt-4o-mini	Fast/cheap model for classification
EMBEDDING_MODEL	No	text-embedding-3-small	Model for failure memory embeddings
API_PORT	No	8000	FastAPI server port
API_HOST	No	0.0.0.0	API host
NEXT_PUBLIC_API_URL	No	http://localhost:8000	Backend API URL for frontend
MAX_RETRIES	No	5	Maximum retry attempts per task
MAX_BUDGET_USD	No	1.00	Maximum cost budget per task
CIRCUIT_BREAKER_THRESHOLD	No	3	Consecutive failures before circuit break

See .env.example for a copy-paste template.

---

Layer	Technology	Role
Orchestration	OpenAI Agents SDK	REJD loop with specialist agent handoffs
Failure Memory	Redis Agent Memory Server	3-tier: working memory, long-term failure graph, episodic traces
Execution Sandbox	Blaxel Firecracker microVMs	Perpetual state, 25ms resume, checkpoint/restore
API	FastAPI + SSE	Task management, memory search, real-time streaming
Dashboard	Next.js + Tailwind + shadcn/ui	Timeline, failure sidebar, cost tracker

---

Contributing

We welcome contributions. See CONTRIBUTING.md for:

• Development environment setup
• Code style requirements (ruff, mypy)
• PR process and review checklist
• Architecture overview for new contributors

---

License

Apache 2.0 -- see LICENSE for the full text.