ReloLayer

Personal AI, outside the model

Change the model.
Keep your AI.

ReloLayer is a small experiment in building personal AI you can inspect and move: a designed identity in plain files, a replaceable model underneath, and fewer reasons to start over when providers change.

Get the starter kit See the experiment
Prototype 001 4 identity files DeepSeek ↔ Kimi plain Python harness
~/my-personal-ai 
identity/
├── persona.md
├── voice.md
├── boundaries.md
└── memory.md

$ export PERSONAL_AI_MODEL=deepseek/...
✓ identity loaded

$ export PERSONAL_AI_MODEL=kimi/...
✓ same identity package

model = replaceable instrument

The problem we ran into

The model was the replaceable part.

Building distinct personal assistants taught us that the difficult part was not calling an LLM. It was keeping the important layer coherent when the LLM changed: identity, memory rules, boundaries, permissions, and the accumulated logic around the person.

01 / identity

Readable, not hidden.

The starter begins with four Markdown files you can inspect, edit, diff, back up, and move.

02 / models

Replaceable underneath.

Point the same identity package at another compatible model and see what survives, what drifts, and what actually depended on the provider.

03 / control

Your rules stay visible.

Boundaries and memory rules live beside the identity instead of disappearing into a vendor-specific black box.

04 / reality

No magic-memory claims.

The v0.1 starter uses manual durable memory. Automatic long-term memory, agent skills, permissions, and routing are work still to be proven.

How the starter works

Keep the identity above the model.

The starter is intentionally boring technology. That is a feature. Plain files go in; a tiny harness assembles them as context; an OpenAI-compatible endpoint does the inference.

PERSONAL_AI_MODEL=provider/model-name

Change one environment variable. Keep the identity files. Start a fresh comparison.

Prototype 001

One designed identity. Two underlying models.

The first public artifact shows the actual files, the rules they contain, a live reference assistant, and matched DeepSeek/Kimi responses. The point is not that portability is perfect. It is that the identity can be made explicit enough to test instead of being trapped inside one chat product.

  • Real files on disk
  • Readable boundaries and identity rules
  • Matched prompts across two models
  • Limitations stated instead of papered over
▶ Prototype 001

2m40s explainer

YouTube embed drops here after publication.

Free beta artifact

Build the smallest version first.

ReloLayer Starter Kit v0.2 is deliberately tiny: four identity files, a README, and an optional no-dependency Python client. Start with no code at all: customize the files with the AI assistant you already use, upload them back for a quick persona test, then use the harness when you want a cleaner model-to-model comparison.

Beta access + occasional ReloLayer project updates. No list selling, no mailing-list circus.

persona.mdidentity & posture
voice.mdtone & interaction
boundaries.mdtruth, privacy & action rules
memory.mdmanual durable memory
chat.pytiny compatible-endpoint harness
Not included yet: automatic durable memory, tool permissions, routing, background autonomy, or a production UI.

Help choose what gets built next

What would make this useful to you?

We are testing whether technically curious builders actually want a cleaner way to create and own a personal AI. This is not a newsletter-growth exercise. Tell us what you are trying to make.

Your response is stored for this ReloLayer market test. We do not sell the list. You can also write directly to relolayer@gmail.com.

What ReloLayer is

A lightweight, user-controlled layer for building a personal AI whose designed identity, memory rules, permissions, and skills can remain coherent while underlying models and services change.

What it is not

Not a foundation model, not another generic chatbot subscription, not a claim that cross-model identity is perfect, and not yet a finished memory or agent platform.