Notes, documents, decisions and years of AI conversations hold what your organisation knows, scattered where nobody finds it again. MemPalace turns that material into an organised, searchable memory for your own AI tools, on a machine you control. Every record can be traced to its source and deleted on request.
MemPalace stores text verbatim and never summarises it away. My pipeline decides what goes in, how it is organised, and how it is found again.
Chat histories, notes, documents and transcripts are cleaned into one format. Claude, Google AI Studio, Gemini and HTML transcripts work out of the box; other sources are added for your project. Noise is stripped out, the content itself stays intact.
Each conversation is split where the subject actually changes. A local language model labels every topic with what you were trying to do and why, and files it under one of your own wings.
Search runs on a specialised embedding model. A custom adapter can be trained on material you provide, or on generic artificial data in your field, and is always tested on your material inside your environment. Personal data never goes into training, so nothing about the people in your records ends up in the model.
The palace answers through MCP, so Claude Desktop, Claude Code or a local model on your own GPU can search your history and quote it back, with the source of every hit.
Two things make AI memory hard under GDPR: sending data to outside processors, and models that absorb personal data and cannot forget it. MemPalace avoids both. Processing stays on your premises, and personal data lives only in records that can be deleted one by one, never in model weights.
Compliance itself depends on how the system is used in your organisation: what you store, who can access it, and how long you keep it. I can help you set it up with that in mind and document how the data flows.
One thing to know: if you connect the palace to a cloud assistant such as Claude, the search results that assistant asks for are sent to it. For fully offline use, pair it with a local model.
The working reference system is built from two years of my own AI conversations: 28,829 topics, stored as 36,885 memory pieces across 7 wings. The test uses 300 cases drawn from conversations the model never saw in training, searched against the whole palace.
| Measure | Base bge-large | MemPalace adapter |
|---|---|---|
| Right result ranked first | 8% | 65% |
| Right result in top 5 | 19% | 85% |
| Right result in top 10 | 23% | 89% |
These numbers come from my own corpus and my own held-out test, not a public benchmark; results on your data will differ, which is why every build includes the same kind of test on your material, run inside your environment, without training on it. MemPalace itself is an open-source project (MIT licence) by its original authors. The import and labelling pipeline, the training setup and the trained adapters are my own work.
Every engagement starts with a short call about your data, your hardware and what you want the memory to do.
Pricing on request
Pricing on request
Hourly or per project
A computer with a modern NVIDIA graphics card, running Linux or Windows with WSL2. The reference system runs on an RTX 5090. How much you need depends mostly on how much history you have, so we check your setup before anything is installed.
No. Chat history is where it started, but the pipeline works on text, so notes, documents, transcripts and other written material can go in too. Out of the box it reads Claude, Google AI Studio and Gemini exports and HTML transcripts; other formats are added as part of your build.
Not during import, labelling, training or storage: all of it runs locally. The one exception is your choice of assistant. If you query the palace from a cloud assistant such as Claude, the results it retrieves are sent to that assistant. Use a local model if nothing may leave the machine.
Not your personal data, ever. A custom adapter can be trained on material you provide, such as routines, guidelines and other documents without personal data, or on generic artificial data written for your field. It is then tested on your real material so we know it works. Keeping personal data out of training is a design choice: once it is trained into a model, it cannot be reliably removed again.
Their records are removed from the palace, and the search model needs no change, because it never contained them. Source files and backups are handled by your normal routines, which we set up together.
Yes. New material goes through the same pipeline and is added to the existing palace as new records. Nothing needs to be retrained.
Tell me what you have and what you want to find again. I will tell you honestly whether MemPalace is the right fit.
