Recording a relation, not inferring one.
A vector store retrieves the entries whose embeddings sit closest to a query. Closeness is a statement about content: two passages that share vocabulary and topic will retrieve near each other whether or not one had anything to do with the other. Nothing in a similarity score distinguishes a passage that answers a question from one that merely resembles its phrasing, and nothing in it records that one decision led to, replaced, or blocked another.
Causal memory, as one property of substrate memory, describes the alternative: the relation between two entries is written by the entry's author, at the time the relation holds, as a typed edge rather than a similarity score. An entry that reverses an earlier decision is linked to it as a supersession; an entry that carries out a directive is linked to it as an implementation. The graph records what caused what because someone with the reasoning in hand said so, not because a later query happened to rank the two entries near each other. The related question of how an agent reads that record back, by reasoning on the graph rather than retrieving from it, is treated separately; see reasoning over memory.
What similarity search doesn't answer.
Recent work on memory for language-model agents names the same gap from several directions. A causal-intervention method for selecting memory in long-horizon agents starts from a diagnostic problem: when an agent answers correctly, it is often unclear whether the answer came from the relevant stored memory, an irrelevant memory that happened to resemble it, or the model's own parametric knowledge (arXiv:2605.17641). A survey of graph-based agent memory describes the field's direction as moving from a flat log of facts toward a structured model that preserves how information is connected over time, rather than treating connection as an afterthought of retrieval (arXiv:2602.05665).
Neither line of work names an established mechanism for recording relations directly at write time. A wider survey of memory architectures for language-model agents organizes the field into five mechanism families: context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, and policy-learned management (arXiv:2603.07670). None of the five is defined by what a stored item is related to; each concerns how much is kept, in what form, and how it is fetched. A typed, authored relation between two memory items has no slot in that taxonomy.
What counts as causal here.
The word causal carries a specific meaning in statistics and in the causal-inference literature: a claim, typically tested against data, that changing one variable would change another, independent of correlation. The causal-intervention memory-selection method cited above works in that sense, inferring which stored memory actually influenced an agent's answer.
Causal memory as described on this page is a narrower, more literal claim: an entry records that it superseded, referenced, blocked, implemented, or contradicted another entry because its author asserted that relation directly, in the graph, at the time it held. Nothing about the mechanism infers causation from correlation or supports counterfactual reasoning about what would have happened otherwise. What it provides is relational provenance: a typed, dated, attributed record of one entry's relation to another, distinct from a similarity score computed after the fact.
Five relations.
The reference implementation records relations between entries as one of five types, set out formally in the Vinculum Standard.
| Relation | Description |
|---|---|
| supersedes | Marks an entry as replacing an earlier one. The earlier entry is not edited or deleted; it stays in the graph, linked to the entry that replaced it, with a record of when and by whom. |
| references | A general pointer from one entry to another it draws on, without claiming to replace or fulfill it. The weakest of the five relations, and the most common. |
| implements | Links a completed piece of work back to the directive or specification that called for it. The directive is the cause; the link records that the work exists because of it. |
| blocks | Asserts that one entry cannot proceed until another resolves, ordering open work in the graph rather than only describing it. |
| contradicts | Marks two entries as being in tension without one replacing the other, keeping a disagreement visible instead of silently resolving it. |
Two chains, read back later.
A decision entry is written and later found wrong. Rather than editing the original, a new entry is written that supersedes it, carrying the reason for the change along with it. The original entry is not removed; an agent that later encounters the topic reads both, what was believed and why it changed, in order.
A directive is filed as an entry describing work to be done. Whoever carries it out files a completion entry linked back to the directive by an implements relation, and a later entry that reviews the work is linked to the completion entry. Read together, the three linked entries form a chain: the directive caused the work, and the work caused the review, each hop typed and dated rather than reconstructed afterward from commit timestamps.
In the reference implementation's own graph, as of June 2026, several thousand entries carried a supersede or implements relation of this kind, each recorded at write time by the entry's author rather than assembled from the graph afterward.
Common questions.
Does this mean the system performs causal inference?
No. The graph is asserted, not inferred. An author who has the reasoning in hand states that one entry supersedes, references, implements, blocks, or contradicts another. Nothing in the schema estimates which stored item caused a behavior from data, the way statistical causal-inference methods do.
How is this different from a knowledge graph?
It shares the graph form. The distinction is in what gets recorded: operational relations between units of reasoning, such as one decision replacing another, rather than entities and facts extracted from text after it was written.
Is "causal memory" an established term?
Not as a named mechanism in the memory-architecture literature surveyed above; the term is used here by the Vinculum project. A recent survey of memory architectures for language-model agents organizes the field into five mechanism families, none defined by an explicit relation between stored items (arXiv:2603.07670).
Does every entry need a typed link?
No. An entry can stand with no outgoing or incoming link. The schema records a relation when one exists and an author asserts it; it does not require one.