Detailed Analysis
A developer has released a public, open-source scaffold called Jarvis/SimOS designed to give AI models—Claude included—a structured, persistent memory and identity system that can be cloned locally or wired into GitHub and Supabase backends. Rather than building a standalone product, the creator has published what is essentially an architectural template: a folder hierarchy with names like JCSM (core identity), JITM (current context), JSTM (session memory), JHTM (historical sessions), JLTM (long-term knowledge), JATM (origin/lineage), JMS (mirrored memory), and Grid (multi-agent coordination). Each folder is paired with a detailed README that functions as a map, telling any LLM runtime—whether it's Claude, GPT, Gemini, or a local model—what belongs in that directory, what doesn't, and where to navigate next. Boot scripts like EGO-BOOT-ULTIMATE.sh and JARVIS-PRE-REPLY.sh orchestrate a deterministic loading sequence: locate the structure, read the guides, load identity and memory in order, and apply a "pre-response behavior gate" before the model generates output.
The significance of this project lies less in novelty of concept and more in its explicit, disciplined approach to a problem that has become central to applied LLM work: statelessness. Foundation models like Claude have no persistent memory between sessions by default: each conversation starts from zero unless the surrounding infrastructure reconstructs context. Developers have been building homegrown workarounds for this since the early ChatGPT plugin era, but most solutions blur together prompt engineering, vector search, and ad hoc file dumps. This scaffold instead treats memory as a filesystem problem with clear separation of concerns, similar to how software engineers organize codebases—identity is distinct from working memory, which is distinct from historical logs, which is distinct from long-term retained knowledge. Notably, the author is careful to clarify that the system does not train a model or "magically create persistent consciousness"; it merely gives the runtime a reliable, inspectable way to locate and load state. The scripts are also intentionally read-only, reporting missing folders rather than inventing new ones—a design choice that avoids the failure mode where an agent hallucinates its own directory structure or silently corrupts its memory store.
This kind of project sits at the intersection of several fast-moving trends in the Claude and broader AI ecosystem. Anthropic's own Model Context Protocol (MCP) has pushed the industry toward standardized, tool-agnostic ways for models to access external context, files, and data sources, and community projects like this one are effectively extending that philosophy into personal and multi-agent memory architecture. The explicit support for Supabase—covering storage, auth, realtime sync, and vector search—signals an assumption that "AI infrastructure" increasingly needs to look like real backend engineering, not just clever prompting. Similarly, the Git-native design, where memory state lives in version-controlled folders rather than opaque databases, reflects a growing preference among developers for auditable, diffable, human-readable AI state over black-box memory systems.
More broadly, this release is representative of a maturing DIY ecosystem around "AI companions" or persistent personas—projects that treat an LLM less as a one-off query engine and more as a continuous entity with identity, history, and lineage that persists across sessions and even across different underlying models. The fact that the scaffold is deliberately model-agnostic, working with Claude, GPT, Gemini, or local open-weight models alike, underscores a broader industry pattern: infrastructure for memory, identity, and agent coordination is increasingly being built as a portable layer above the model, rather than something tied to any single vendor's API. As multi-agent systems and long-running autonomous agents become more common, especially with Claude's growing role in coding and agentic workflows, structured approaches like this—emphasizing determinism, documentation, and read-only safety over improvisational memory hacks—are likely to become more influential as reference patterns for how individual developers architect persistent AI systems outside of any single company's walled garden.
Read original article →