What is Alma? Features, Pricing & Tutorial for AI Agents (2026)

The Problem Alma Solves

Modern AI agents are often treated as independent silos, each building its own shallow understanding of your preferences, working style, and values. You end up repeating the same instructions—"keep responses concise," "always use TypeScript," or "I prefer academic tones"—every time you open a new chat session. This fragmentation turns your AI interactions into repetitive administrative tasks rather than productive collaborations.

Furthermore, standard AI platforms often hoard this "memory" behind their own closed-source walls. You have no reliable way to move your identity or preferences from one tool to another. Beyond the portability issue, there is the persistent security concern: allowing an LLM to freely read and write to your long-term memory is a massive leap of faith that most users aren't equipped to manage safely.

Alma addresses these pain points by inserting a local, transparent layer between you and your AI agents. By utilizing the Model Context Protocol (MCP), it forces agents to request access to your "self-model" and requires your explicit, human-in-the-loop approval before any new information is stored. You are no longer just a user; you are the gatekeeper of your digital context.

In this tutorial, you will learn exactly how to use Alma—step by step—to regain control over your AI memory.

How to Get Started with Alma in 5 Minutes

1. Prerequisites: Ensure you have Rust and Cargo installed on your system, as Alma is currently a source-built project.
2. Build the Project: Clone the Alma repository, navigate to the folder, and run cargo build --release to generate the necessary binaries.
3. Quickstart Setup: Run ./target/release/alma quickstart in your terminal to initialize your local ~/.alma/alma.db store.
4. Connect Your Agent: Use the CLI to hook into your existing AI tools, such as ./target/release/alma connect cursor --apply or ./target/release/alma connect claude --apply.
5. Reload and Verify: Restart your AI IDE or agent client; it should now detect the Alma MCP server and begin requesting contextual data as needed.

How to Use Alma: Complete Tutorial

Managing Your Self-Model with the TUI

The alma-companion utility is the primary interface for users who prefer a visual overview over raw JSON editing. Once you have your environment set up, run alma-companion to view your existing facets. This tool allows you to browse through your stored preferences, such as work-style settings or identity markers, without needing to interact with the database directly.

Within this interface, you can preview the "Reading" that an agent would receive if it queried your memory. This is critical for transparency; it allows you to see exactly what you are exposing to an AI before a connection is even established. You can also review any pending proposals here, ensuring you never accidentally authorize a memory write you haven't personally vetted.

Handling Memory Proposals and Consent

Alma operates on a strict "no-trust" architecture. When an AI agent determines it has gathered enough evidence to form a new fact about you, it calls alma_propose_facet. At this point, the write is not durable; it is merely a suggestion. You must move to your CLI or the Companion TUI to approve the request.

Once you approve, a one-time token is generated, allowing the agent to call alma_record_facet. This two-step dance ensures that your memory remains an append-only event log that you fully control. If an agent tries to modify a core preference without your permission, the system will effectively ignore the attempt, keeping your self-model clean and unpolluted by erroneous AI assumptions.

Exporting and Porting Your Memory

The most powerful feature of Alma is its commitment to vendor neutrality. Because your data is stored in a local SQLite file and audited through an event log, you are not trapped. When you need to move your identity to a new system, use the alma_export_bundle command.

This generates a JWS-signed file using Ed25519, which acts as a portable, cryptographically verifiable representation of your self-model. Because the public key is embedded in the header, you can verify your data integrity without a central authority or a complex key-exchange process. Simply import this bundle into a new Alma installation to re-initialize your preferences exactly as they were.

Alma: Pros & Cons

Pros	Cons
Full data ownership and local privacy.	Experimental project; APIs may shift.
Vendor-agnostic portability via JWS.	Requires manual compilation from source.
Granular, human-in-the-loop consent.	Higher technical barrier to entry.
Transparent and audit-able memory logs.	No pre-packaged binaries for easy install.

Alma Pricing: Free vs Paid

Alma is an open-source project and is currently entirely free to use. There is no "pro" tier or gated functionality. Because it is a local-first, self-hosted tool, the development team has opted for an open-source distribution model rather than a SaaS subscription.

Since the project is in an experimental hobbyist state, there is no infrastructure-heavy cost for the maintainers to pass on to you. You are responsible for your own data storage and compute, meaning the "cost" is simply the time you spend maintaining the installation and managing your memory facets.

👉 Check the latest pricing on the official Alma website.

Who is Alma Best For?

For Power Users: If you use multiple AI assistants throughout the day, Alma allows you to maintain a consistent baseline of your coding style and personal preferences across all of them. It effectively ends the frustration of "training" each new chat window from scratch.

For Developers: If you value transparent, auditable infrastructure, Alma provides a clean, local-first way to manage state. The use of SQLite and an append-only event log provides a familiar, inspectable data model that you can verify at any time.

For Privacy Enthusiasts: If you are uncomfortable with the idea of a third-party AI provider "remembering" your personal habits, Alma puts you back in charge. By forcing every single memory write to go through a human-approval process, you ensure that no unauthorized data is ever stored on a remote server.

Alternatives to Alma

While there are other ways to manage AI context, most exist as proprietary features within specific platforms, such as OpenAI's "Memory" feature or Claude's "Projects." These alternatives are easy to use but ultimately lock your data into a specific vendor's ecosystem. Other open-source local LLM managers like Ollama offer ways to run local models, but they rarely provide a structured, universal "self-model" layer like Alma. Alma stands out because it focuses specifically on the interoperability and governance of your personal information, rather than just the storage of model weights or local chat logs.

Final Verdict: Is Alma Worth It?

If you are a technical user who is tired of the vendor lock-in associated with AI memory and you value data sovereignty above convenience, Alma is an excellent project to adopt. It is not currently for the casual user who wants a one-click installer, but for those willing to build from source, it provides a powerful, transparent, and portable foundation for their AI workflows.

📋 Disclosure: This is an independent tutorial based on Alma's publicly available documentation and website content as of June 23, 2026. GitNeural is not affiliated with, sponsored by, or endorsed by Alma or github.com. Pricing and features may have changed — always verify on the official Alma website.