Idea: Vestige, a local-first memory MCP server that pairs with Planner Agent V3

[samvallad33]

Hi Juan Jose, thanks for this collection. Planner Agent V3 with full MCP support is the piece that got me thinking about posting here.

I am Sam (github samvallad33), and I build Vestige, so this is my own project. I am not proposing a PR into your repo, I just think it fits a gap and wanted to float it as an idea for you and the community to react to.

The gap I noticed: the collection has a lot of capability but no dedicated long-term memory tool. The memory that exists today is Open WebUI's native user memory surfaced inside Planner Agent V3. That is useful, but it is plain storage, not retention-aware recall.

What Vestige is, plainly:

Vestige - Local-first cognitive memory MCP server for AI coding agents. FSRS-6 retention, prediction-error gating, active forgetting, spreading activation, 3D dashboard. Single Rust binary. npm install -g vestige-mcp-server.

Why I think it is complementary rather than a replacement: Planner Agent V3 already speaks MCP, so Vestige is something your existing pipe can connect to instead of a new framework to adopt. It stays local-first and user-owned, so nothing leaves the machine, which matches the self-hosted spirit of the rest of the collection.

Honest fit notes so I am not overselling:

• Vestige exposes itself over stdio MCP. The clean path into Open WebUI is either Planner Agent V3's own MCP client or mcpo, the Open WebUI MCP-to-OpenAPI proxy. I have not tested the exact wiring against your Planner build, so I would treat that as the open question.
• It was built for coding agents first. The retention and forgetting model is general, but the dashboard and some tooling assume an agent workflow.

Rough shape of how mcpo would expose it, for reference:

npm install -g vestige-mcp-server

then run mcpo in front of the stdio server

mcpo --port 8000 -- vestige-mcp-server

register http://localhost:8000 as an OpenAPI tool server in Open WebUI

Repo and full docs are under github samvallad33 / vestige if you want to look closer. Totally fine if this is out of scope, I mostly wanted to put it in front of the right person. Happy to answer anything.

---

Hola Juan Jose, gracias por esta coleccion. Planner Agent V3 con soporte completo de MCP fue lo que me animo a escribir aqui.

Soy Sam (github samvallad33) y soy quien desarrolla Vestige, asi que es mi propio proyecto. No estoy proponiendo un PR a tu repo, solo creo que encaja en un hueco y queria dejarlo como idea para que vos y la comunidad opinen.

El hueco que note: la coleccion tiene mucha capacidad pero no hay una herramienta dedicada de memoria a largo plazo. La memoria que existe hoy es la memoria nativa de Open WebUI que aparece dentro de Planner Agent V3. Sirve, pero es almacenamiento simple, no recuerdo con retencion.

Que es Vestige, en simple:

Vestige - Servidor MCP de memoria cognitiva local-first para agentes de IA. Retencion FSRS-6, gating por error de prediccion, olvido activo, activacion por difusion, dashboard 3D. Un solo binario en Rust. npm install -g vestige-mcp-server.

Por que creo que es complementario y no un reemplazo: Planner Agent V3 ya habla MCP, asi que Vestige es algo a lo que tu pipe actual se puede conectar, no un framework nuevo para adoptar. Es local-first y los datos son del usuario, nada sale de la maquina, en linea con el espiritu self-hosted del resto de la coleccion.

Notas honestas para no exagerar:

• Vestige se expone por MCP stdio. El camino limpio hacia Open WebUI es el cliente MCP del propio Planner Agent V3 o mcpo, el proxy MCP a OpenAPI de Open WebUI. No probe el cableado exacto contra tu build de Planner, asi que lo dejaria como la pregunta abierta.
• Se penso primero para agentes de codigo. El modelo de retencion y olvido es general, pero el dashboard y algunas herramientas asumen un flujo de agente.

Forma aproximada de como mcpo lo expondria, como referencia:

npm install -g vestige-mcp-server

luego corres mcpo delante del servidor stdio

mcpo --port 8000 -- vestige-mcp-server

registras http://localhost:8000 como tool server OpenAPI en Open WebUI

El repo y la documentacion completa estan en github samvallad33 / vestige por si queres mirar mas. Si queda fuera de alcance no hay problema, sobre todo queria ponerlo frente a la persona indicada. Quedo atento a cualquier duda.

[Haervwe]

Hey!! i like the idea! but what i think can be more interesting to do, and can play well with planner and any other agent is using that mcp server as a filter pair so memory actually runs in a separte "mini agent" pre and post completion. !

[Haervwe]

well i started reviwing a little, we can do both a custom tool that inside spawns the rust binary using python , AND and automatic filter logic to retrieve and store memories its a pretty cool project! definetly will be playing with it!