--- title: "A memory that only collects is a haystack" description: "Why we built our own AI memory: search, weighting, contradictions and the rules for maintenance and retention." author: "Matthias Meyer" published: 2026-05-12 updated: 2026-09-30 updates: - date: 2026-09-30 note: "Preserved development context; clarified validity, client support and import limits." language: en tags: ["maschinenraum", "ki-gedaechtnis", "mcp", "eigenprojekte", "architektur"] canonical: "https://studiomeyer.io/en/blog/memory-warum-wir-es-selbst-gebaut-haben" markdown_versions: ["https://studiomeyer.io/de/blog/memory-warum-wir-es-selbst-gebaut-haben.md", "https://studiomeyer.io/en/blog/memory-warum-wir-es-selbst-gebaut-haben.md", "https://studiomeyer.io/es/blog/memory-warum-wir-es-selbst-gebaut-haben.md"] publisher: "StudioMeyer, https://studiomeyer.io (llms.txt: https://studiomeyer.io/llms.txt)" --- # A memory that only collects is a haystack Everyone who works seriously with an AI assistant reaches the same point. You explain the same context for the fourth time. You paste the same backstory into the window for the third time. And eventually you do the job yourself again, because explaining takes longer than doing. That is not user error. A chat without configured access to earlier conversations or stored context does not reliably carry that context into the next conversation. Built-in memory helps but does not meet every company knowledge requirement. The obvious fix is to store everything. That is exactly where the real problem starts. ## Storing Is The Easy Part A system that records every conversation has a mountain of entries after half a year. When somebody then asks "what was the story with the quote for the boating school", there are two bad answers. One is to return everything containing the word "quote". Then a pile of hits sits there and the human searches manually again. The other is to return three of them arbitrarily. Then the decisive one may be missing, and you do not notice, because you cannot know what is absent. A memory that only collects is therefore not a memory but a haystack with timestamps. The value does not sit in the storing, it sits in the question of what comes back. ## Two Kinds Of Memory You Must Not Mix The most important distinction is one you know from psychology and overlook first when building. There are memories of events. On Tuesday the sync ran and this came out of it. And there is knowledge about states of affairs. This company works in Spanish, this client does not want calls before ten. For states of affairs, establish which statement applies to which period and scope. A newer entry replaces an older one only when source and meaning justify it. An earlier price may still matter for an old invoice. For events the opposite holds. Two similar-sounding events are still two events. Merging them because the text resembles itself deletes history. Getting both rules into a system that tidies up automatically was the most laborious part of the whole job. And it is the part where you can tell whether somebody built a memory or labelled a database with a search box. ## Contradictions Are Normal, Not The Exception Across months every system accumulates statements that contradict each other. Not because anybody lies, but because things change and nobody retracts the old note. A memory has to handle that in three stages: notice that two entries conflict. Decide whether one replaces the other or whether both hold side by side. And when in doubt, delete nothing and flag instead. The third stage matters most. A system that deletes under uncertainty loses precisely the cases you would have needed it for. ## What Happens On Retrieval A search in memory is not a text search. It combines several routes at once: the literal match, the fuzzy similarity for typos and word forms, and the semantic proximity, so that "the thing with the certificate" also finds the entry that says "SSL renewal". On top comes time. Newer weighs more than older, but not absolutely, otherwise the decision from a year ago that still stands disappears. And there is a knowledge graph alongside: people, companies, projects and their relationships as a net of their own. That lets us look up explicitly stored relationships, for example "what else is connected to this". ## The Tidying Happens At Night The memory works when nobody is watching. At night a chain runs: condense what was said repeatedly. Score what proved useful and what was never retrieved again. Look for contradictions. Form summaries. The effect is the same as for a human who sleeps. Not more information the next day, but better sorted information. ## Where It Runs The system runs in three places: internally with us, at clients inside their own setups, and as a service you can subscribe to. Access runs over an open protocol that suitable clients such as Claude, Cursor or Codex can support. Each integration and its available permissions require individual checking. Existing conversation exports can supply material for an import. They provide a selected initial collection rather than a guaranteed complete, lossless migration. Confidential contents and the resulting records need review. What it costs is on the product page and not here, because prices change and a blog post does not notice. ## What To Work On First Improving search feels productive, because you see the result immediately. Tidying feels like housework. In reality the tidying decides whether the search is still worth anything a year later. Anyone building such a system should decide in the first week what gets deleted and when. Not once the first search noticeably stalls. This post describes our development work from May 2026. Update of 30 September 2026: providers have developed their memory and import features further. Our own memory also needs maintenance. Automatic consolidation and contradiction detection can make mistakes, so sources, review and purpose-based retention and deletion remain part of operation. ## Read more, Engine Room - [Every Business Gets Its Own AI Agent](https://studiomeyer.io/en/blog/meetmyagent-jeder-hat-einen-agenten.md) - [A Learning Platform Run By Agents](https://studiomeyer.io/en/blog/academy-von-agenten-betrieben.md) - [What We Actually Work With](https://studiomeyer.io/en/blog/arbeitssystem-editor-modelle-werkzeuge.md)