Gentle-AI v4.4.0 Enhances AI Coding Agents with Memory and Workflow

What changed

Gentle-AI has released version 4.4.0, an open-source project that acts as a deterministic engineering environment for AI coding agents. The core functionality of Gentle-AI is to augment existing agents, such as Claude Code, Cursor, OpenCode, Codex, and Pi, by providing them with persistent memory, structured workflows, and evidence generation capabilities. This aims to address the common issue where AI agents forget context between sessions, leading to repetitive tasks and a lack of project understanding.

Key features introduced or enhanced in this release include:

Engram™ — Persistent Memory

This feature allows AI agents to retain project context by writing down what they learn during a session and recalling it in subsequent interactions. This prevents the need for developers to re-explain project details repeatedly, accumulating context over time.

Organic Driven Development (ODD)

ODD focuses on managing work by understanding, risk, and resumability rather than just file count. Small changes are handled inline, while larger tasks are managed through recoverable feature documents. This ensures that work remains lightweight and context is preserved, allowing for seamless resumption of tasks across sessions.

Delegation Logic

Gentle-AI delegates tasks to subagents only when a clear reason exists, such as exploration needs, parallel writing requirements, context backstopping for long sessions, or independent verification for high-risk changes. This prevents unnecessary overhead and ensures efficient use of subagents.

Measured Routing and Benchmarks

The routing rules are informed by benchmarks, demonstrating improvements in efficiency. For instance, a 3-line bug fix saw a reduction in time and cost from 335s/$0.339 to 77s/$0.105 by running inline without subagents. A bug rooted in shared logic improved from 479s/$0.361 to 276s/$0.207 with one explorer. Data migration, previously requiring a feedback round, was accepted on the first round with a different routing strategy. Benchmarks also indicate that Gentle Shell, a specific Pi integration, was accepted on the first round in all tested scenarios against Codex CLI, though large tasks still take longer.

Test-First by Default

When a relevant runnable test and a clear expected outcome exist, Gentle-AI applies test-first development. This involves capturing a failing behavior test before implementation, making it pass, and then refactoring while tests remain green. This ensures that each requested rule is proven and existing functionality is preserved.

Receipt-Driven Development (RDD)

RDD is enabled by default to ensure that reviews are based on a frozen candidate, guaranteeing that the evidence belongs to the exact version being reviewed. Reviews are proportionate to the risk, with agents naming specific risk lenses (4R) that fit the touched code.

Deterministic Design

The gentle-ai binary aims for deterministic behavior, ensuring that the next valid step is predictable, unlike models that might guess differently each time. This provides a consistent workflow rather than mere suggestions.

Integrations

Gentle-AI supports 17 agents, including Pi, OpenCode, Claude Code, Codex, and Cursor, leveraging their native capabilities. Gentle Shell is highlighted as a complete workspace for Pi, bundling Pi, Gentle AI, and Engram for a self-contained experience.

Why it matters for builders

Gentle-AI's approach directly addresses the limitations of stateless AI coding agents, which often require developers to re-establish context and project history. By introducing persistent memory (Engram™) and structured workflows (ODD), builders can leverage their AI agents more effectively, reducing development friction and improving the overall quality and consistency of AI-assisted code. The focus on deterministic behavior and evidence-based reviews also enhances trust and reliability in AI-generated outputs.

Practical impact

Developers using AI coding agents can integrate Gentle-AI to gain persistent memory and structured workflows. This means less time spent re-explaining project context and more time on actual development. Builders can explore the 17 supported agent integrations, starting with their current tools like Claude Code, Cursor, or Codex. The project's open-source nature and lack of agent lock-in allow for flexible adoption. The detailed benchmarks provide concrete examples of efficiency gains, encouraging builders to test these improvements in their own projects.

Caveats and source limits

The provided source is a GitHub repository description and excerpt, offering detailed feature explanations but lacking independent benchmark results or comparative performance data beyond the project's own benchmarks. Specific pricing for any potential enterprise features or support is not mentioned, as the project emphasizes being donation-supported and free of paywalls. The exact release date for version 4.4.0 is not specified beyond the general publication date of the repository information. The performance gap for large tasks mentioned in the benchmarks (approximately three times longer for Gentle Shell vs. Codex CLI) indicates areas for future development.