RameSystem connects agents, models, tools and verifiable memory to create development experiences that are continuous, reviewable and recoverable.
Conceptual illustration of the interface.
Each piece has a clear role. What is already in development is kept separate from the research lines.
Not a RameSystem product.
RAME Code is a separate open-source project, newly implemented by an individual developer during the Meta Global AI Developer Hackathon 2026. It draws conceptual inspiration from the earlier RAME research, but it is not a release or fork of RameSystem and does not reuse or depend on its code.
ramecode.comThe engine that turns a request into controlled work.
Coordinates models, files, terminal, external tools, permissions and results. Its modular architecture lets you use different providers and local models without redesigning the whole experience.
In developmentPersistent memory for decisions that must outlive a conversation.
Rame Memory is an external service for versioned, auditable memory. It keeps verified decisions, their provenance and their history. A model can query the relevant information, but it cannot silently change what the user authorized.
Experimentally validated prototypeResearch into our own neural architecture.
RAME Core 2 studies a neural core with relational state, memory and incremental processing. It stays separate from the commercial products until it shows enough capability and stability.
Experimental researchRame Memory is the external memory and RAME Core 2 is an independent research line: they are not the same thing and are not presented as if they were.
Before any product there was research. This is what was tried, what worked and what did not.
A recurrent core different from the Transformer, with fixed state, relational memory and sparse selection. It learned some tasks, but its original core was frozen because it did not generalise enough. RAME-Lab was the lab for isolated primitives (vector memory, pointers, routers, scaling): it produced useful results and many limits, and no finished conversational brain.
Research · core frozenThe later attempt: its own recurrent architecture, long context with little memory growth, real neural pretraining, progressive growth and experimental experts. It is still research and does not yet replace any general-purpose model.
Experimental researchCurrent and historical facts with versions, provenance and confirmations; exact restarts, a shadow-learning log and defence against content without authority. It worked conservatively and with confirmation. Autonomously deciding what to store from free language failed.
Prototype with known limitsIn September, before the hackathon, RameSystem ran a real cascade between models (from a small one to larger ones and back to the small one) and kept its state on resume. Still unresolved: tying error detectors to the cascade, deciding when to escalate from calibrated evidence, and showing that returning to the small model keeps quality.
Tested with real models · incompleteEarlier private results are not presented as the work of any later project.
The user talks in an interface; Rame Agent does the work and connects to the chosen model, the memory and the tools.
The model reasons and proposes. RameSystem controls the tools. The user keeps authority over important changes.
Rame Memory keeps explicit decisions with version, provenance and authorization. A later session can retrieve the current version without depending only on the chat history.
A model can query the memory through MCP in read-only mode. Important changes go through a separate control channel that the model cannot authorize by itself.
RameSystem is designed to work through adapters with external providers, compatible APIs and local models. Each integration must be tested and configured specifically.
External providers through their API.
Running on your own machine.
Models hosted on your infrastructure.
Such as RAME Core 2, when it is ready.
These are planned integration categories, not a guarantee of commercial compatibility with any specific provider.
Change models without changing your whole way of working.
Sessions and processes are designed so they can be resumed.
Important changes can be shown before they are applied.
Decisions keep versions, provenance and history.
RameSystem is building the pieces that let agents work for longer, use different models and keep verified decisions without turning chat history into an opaque authority.
If you want to follow the development, try what comes out, or collaborate, write to us. We reply personally.
RameSystem · a technology project by Antoni Caimari Caldes