WHITEPAPER · VERSION 1.0.3 · JUNE 10, 2026
The StudioMeyer Memory whitepaper: how a bi-temporal knowledge graph gives AI agents a memory that knows what is still true today. Architecture, benchmark methodology, an honest competitive comparison and documented limitations — no marketing fog.
PDF · German · approx. 330 KB · no form, no login
*Internal 50q stratified run (S957b, GPT-4o as judge), prior to the v3.16.11 fix — a clean re-run is pending. Methodology and limitations are openly documented in sections 4 and 9 of the whitepaper.
Plainly explained
Eight things StudioMeyer Memory does — explained without jargon. As much as necessary, as simple as possible. The technical depth lives in the whitepaper.
While you work with your AI, Memory captures the essentials: decisions, facts, preferences, connections. Not every word — the things you will still need tomorrow. A built-in doorman filters out the noise before it ever gets stored.
A sticky note does not know that “Anna” is your tax advisor and that “Project Finca” belongs to her. A knowledge web does: every piece of information is a point, and connections show what belongs together. That is why your AI can answer questions that go around several corners.
“The server crashed yesterday” is an experience. “The server runs in Frankfurt” is a fact. Memory treats them differently, just like your brain: experiences may fade over time, facts stay stable — even when they are old. Truth does not expire just because it has been around for a while.
For every fact, Memory records two things: when was it true in the real world, and when did we learn it? Sounds like a detail, but it is the difference between “the customer lives in Berlin” and “the customer lived in Berlin until March, in Hamburg since”. You can ask: what did we know on March 15?
When new information contradicts old information, nothing gets silently overwritten. Memory flags the contradiction and keeps both versions with their timestamps. You keep the history, and the AI does not have to guess which version is right.
You ask about “that certificate thing”, but it was stored as “SSL renewal”. Memory finds it anyway, because several search methods look for meaning in parallel, not just for words. And when nothing reliable is stored, the system says exactly that instead of making something up. Results that prove themselves grow stronger over time.
Between sessions the dream cycle runs: it condenses similar experiences into one stable fact, checks the stock for contradictions and adjusts how much the system trusts individual memories. Whatever has not been needed for a long time moves to the background — it is not deleted. In the morning the memory is tidier than it was at night.
Hosted in Frankfurt, GDPR-compliant, every customer strictly separated from all others. You can export or delete your entire memory at any time — right to be forgotten included. Memory belongs to your toolbox, not the other way round.
Three core claims
Not a product brochure. The whitepaper explains the architecture, discloses the measurement methodology and names the limits — including the cases where a simple Markdown file is all you need.
Static notes like CLAUDE.md, accumulating notes like Auto-Memory, structured knowledge graphs. The market lumps them together. The whitepaper separates the layers cleanly — and says honestly when layer 1 is enough for solo developers, and when you need layer 3: cross-tool, multi-tenant, contradiction detection, audit trail.
Most memory systems overwrite old truths. StudioMeyer Memory stores, per fact, when it was true in the world and when it entered the memory. “What did we know on March 15?” becomes an answerable question — essential for support reconstruction, compliance audits and voice agents.
New since June 2026: a dream cycle condenses episodic memories into stable facts between sessions, checks for contradictions and rebalances confidence. And retrieval stays honest — when nothing reliable is stored, it says so instead of guessing. Results that prove themselves grow stronger over time.
Contents
From the problem definition through architecture and measurement methodology to pricing, roadmap and a dedicated limitations section.
Download
Version 1.0.3, June 10, 2026. In German, as PDF, no form and no login. Read it, hand it to your team, or feed it to your favourite LLM.
Whitepaper as PDFFAQ
A memory for your AI: it remembers what matters, finds it again — and knows what is still true today.
Honest answer: if you only chat with an AI occasionally, a simple notes file is often enough. Memory pays off once several tools, projects or people are involved — or when it matters what was true when.
No. Your AI takes notes while you work — decisions, facts, connections. You can also import existing histories from ChatGPT, Claude or Gemini.
Two time axes per fact: when was something true in the real world, and when did we learn it? If your customer moves from Berlin to Hamburg in March, both states are kept — and “what did we know on March 15?” becomes an answerable question.
Far less often. Memory delivers real stored facts and says openly when nothing reliable is on file — so the AI no longer has to guess. With language models, invention can never be ruled out entirely; the whitepaper says so honestly.
Tidying up, like during sleep: similar experiences are condensed into one stable fact, the stock is checked for contradictions, and trust in individual memories is recalibrated. Nothing gets deleted in the process.
Because our customers speak German. The product itself and this page speak German, English and Spanish — the whitepaper deliberately appears first in the language we consult in.
In the PDF: architecture in section 2, measurement methodology in section 4, limitations in section 9. Our line: as much openness as possible, as little blueprint as necessary.
Memory runs as a hosted MCP server: one URL, magic-link login, free plan forever. The live demo shows a real knowledge graph in 3D.
