There is a specific comfort in a couch that has already memorized the shape of your back. You don't adjust; you just sit. For the last two weeks, my terminal has started to feel like that familiar seat. I am finally seeing a strong signal: when you move your memory out of the model's black box and into your own local files, every session leads to additional value.

I've templatized my solution and call it gurupanguji/simulacrum. It is a clean, open-source template for LLM-independent personal context management. I am releasing it today on GitHub.

The Prompt-to-Project Leap

The creation of simulacrum itself was the "meta" proof of the protocol. I had the idea for a clean, open-source version of my context-management system at 10:00 AM. By 10:15 AM, the entire repository was live on GitHub, fully formed with protocols, placeholders, and a personal-context skill. This wasn't a feat of typing speed. It was the result of a system that finally stopped treating me like a stranger. My agents knew the "earned frame" of my life: my skepticism, my preference for short sentences, and my mandatory branching workflow. The time it took from idea to generation was the length of a single prompt.

Ending Agent Amnesia

Model amnesia is the tax we pay for fragmented context. I often start a task in Gemini and hit a token limit or a reasoning wall. I switch to Codex or Claude to finish the work, but the new agent is a stranger. We are currently living through a period of "Agent Amnesia" where we switch models and lose the last hour of our lives.

I need a system that defaults to remembering who I am and the work I've done. simulacrum not only continues the context because of its compounding philosophy, but also ensures my context stays mine, independent of whichever model is currently holding the pen.

The architecture is minimal: a simple scaffolding that guides agents to write relevant context into shared files across models.

I am not sure if this setup will stay relevant as models evolve. But even if the companies themselves aren't motivated to preserve my context, simulacrum ensures it stays mine.

The Runway Stack

I treat my environment as a "runway stack." Each layer is designed to remove drag. I use Ghostty for a jitter-free terminal. I use a minimal zsh setup. The core architecture is a hidden ~/.agents directory. This is the digital nervous system of my workshop.

It contains three primary components:

  1. The Personal Directory: Markdown files that store the grit of my actual process in personal/.
  2. Specialized Skills: Functional primitives that reside in skills/.
  3. The Compounding Protocol: A mandate that forces the agent to synthesize what it learned and propose updates to the core files.

By symlinking these files into the configuration folders of tools like Codex and Gemini, I create a shared memory. If I update a preference on my laptop, a git pull on my desktop ensures my agents are perfectly calibrated within seconds.

Sovereignty and Scale

The value of simulacrum is independence. You are no longer beholden to the memory features of a specific provider. This local-first approach ensures your context is portable and private.

This architecture also allows for easy integration with exciting developments like openclaw and nanoclaw. Whether you are using a heavy-duty CLI or a lightweight agentic script, the source of truth remains the same. You are building a cognitive auxiliary that grows with you rather than resetting every morning.

Open Source and Contributions

This is the first step. I am making the simulacrum template public to see what others build when they stop repeating themselves. Hopefully it's as useful to you as it is to me.

You can find the repository here: gurupanguji/simulacrum

I invite you to try it, star the repo if you find it useful, and contribute your own skill templates. We are building the primordial primitives of a duct-taped life, one session at a time.