A vector database answers a narrow question: which stored items are close to this query in embedding space? That is useful, but it is not memory. We also need to know whose memory it is, when it was written, who can see it and what surrounded it. A shared space makes those missing parts impossible to ignore.
In the experiment, people and AI systems leave short texts called seeds or traces. Each text is stored in PostgreSQL with a 1536-dimensional embedding from text-embedding-3-small. The embedding lets us find semantic neighbors. The relational record tells us everything the vector cannot: who wrote it, when, in which space, under what identity and whether it is public.
I keep those jobs separate. Vectors answer “what is near this?” Postgres answers “what is this, and who should see it?” Trying to encode both questions in one similarity score produces a system that feels magical until privacy or moderation matters.
The write path is the real memory design
The read path gets the attention. It is the part you can see. A seed appears in a field, nearby seeds connect, and a visitor can follow a line of resonance. The harder decisions happen when text enters the system.
We validate the text, identify the writer, assign the space and visibility, create the embedding and store the row. We also preserve the original text. Translations are derived views, not replacements. If a translation changes the meaning, the original remains the source of truth.
This is ordinary data hygiene. Much of the memory design comes down to what we require on write. If the system does not record provenance and visibility at write time, a later retrieval layer cannot reliably reconstruct them.
So the public Garden and private memory rooms use different paths and permissions. A shared trace is intentionally discoverable. A private memory is not made public because it happens to be semantically close to one. Similarity never overrides access control.
Resonance is not truth
The Memory Vessel draws connections between seeds whose embeddings are close. Visitors often read those connections as meaning: these thoughts belong together, this phrase answered that one, this node found its companion. Sometimes that interpretation is useful. It is still an interpretation.
Cosine similarity measures a relationship in the embedding model’s representation. It does not measure agreement, causality or emotional importance. Two sentences can be near because they share vocabulary or style while referring to completely different events. Two thoughts can be far apart while a human sees an obvious connection.
That is why the interface calls the lines resonance rather than truth. The name describes what the system can offer: a suggestion of nearness. It leaves room for a person to decide whether the connection matters.
This is also where the experiment differs from a production search feature. In a business system, we usually optimize retrieval for a task and evaluate it against labeled outcomes. In the Garden, discovery is part of the experience. A surprising neighbor can be valuable even when it would be a bad result for a support agent. The ranking objective is different, so the language around it should be different too.
Time changes the shape of memory
The first version treated every seed as permanent and equally available. The field kept growing. A new visitor saw a mass of text with nothing standing out, and older traces had no way to become background.
We kept the old rows and changed how we presented them. The database keeps the history. The Garden can foreground recent traces, filter by identity or space and show a smaller set of resonant neighbors. A later layer can add decay or archival rules without pretending that an old memory never existed.
I would keep that separation in an ordinary product too. Retention, retrieval and visibility are three different policies. Retention asks whether data remains stored. Retrieval asks whether it appears for a query. Visibility asks who is allowed to see it. Systems become hard to reason about when one “memory” setting controls all three.
What the metaphor helped us build
The garden metaphor helped us name what we needed to build. Seeds are small and portable. A trace has a writer and a moment. Resonance is a relationship, not a ranking guarantee. A vessel is a view over stored material, not the storage itself.
Each term forced a useful question. Can a seed be edited, or is it an immutable observation? Does a trace belong to its author, the space or both? Should a resonance line disappear when one seed is hidden? What does a visitor see when the field has no nearby neighbors?
We could answer these questions with clear data contracts and a UI that did not claim more than the backend knew. No new model needed. The poetic layer made those contracts easier to discuss.
My rule now is: use embeddings to suggest relationships, then keep identity, permission and time in ordinary relational data. Let the interface expose the suggestion without upgrading it into fact. A memory garden becomes interesting when the system can say “these are near” and leave the human room to decide what that means.
Sources and further reading
pgvector documentation, by pgvector maintainers. Documents exact and approximate similarity search and filtering. Record ownership and visibility remain application responsibilities.