Memory built for the model, not the human.
The model reasons on it. It isn't a log to read, a summary to skim, or a store to search.
This document specifies substrate memory: a typed, append-only graph of a project's decisions that a language model writes to and reasons on across sessions. It defines the entry shape, the core operations, and the epistemic guarantees a conformant implementation must provide, and nothing about how any one product is built on top. The goal is a primitive small enough that anyone can adopt it, and complete enough that a model built on it stops starting from zero.
MUST, MUST NOT, SHOULD, and MAY are used in the sense of RFC 2119. A conformant implementation is one that satisfies every MUST in §8. Everything else is guidance.A record that reasons back.
For as long as people have written work down (the ledger, the lab notebook, the commit log, the wiki), the record has been dead storage. It holds what you wrote and waits for a mind to come read it and do the thinking. Substrate memory inverts that: the mind is already reading. A session writes a decision, and the next session reads it back, connects it, catches where it breaks something older, and picks up the thread.
The distinction decides the shape. A record written for a human to read later drifts toward prose and summary. A record written for a model to reason on now is typed, linked, and append-only. This specification describes the second and sets the first aside.
a log
A log is append-only but untyped and unlinked, a stream you scroll. Substrate memory is a graph you traverse.
a summary
A summary is a gist written once and frozen, stale the moment after. Substrate memory is the record itself, written as the work happens.
a vector store
Retrieval rediscovers relevance on every query. Substrate memory keeps the structure (types, links, supersession), so relevance is already encoded.
Everything else follows from these.
- The reader is the model.Every design decision resolves in favor of what the model needs to reason well, not what a human needs to browse. Where the two conflict, the model wins. That reverses a decade of memory built for the human.
- Append-only. Supersede, never overwrite.A decision that changes is never edited or deleted. The old entry is kept, and a new entry supersedes it, with the link between them preserved. The changed mind is the record, and the lineage between old and new is the product.
- Typed, not prose.Every entry carries a type and structured metadata, so the model reasons on structure instead of re-parsing haze. A decision isn’t a paragraph that happens to contain one. It’s an entry of type decision, and the model acts on the type without re-reading the prose.
The atomic unit.
decision, implementation, question, note, spec are recommended baseline types.supersedes, replies_to, references. An entry with no links is a leaf; the value compounds with the edges.The five operations.
WRITE
Append a typed entry. Never mutates an existing entry; writing is the only way the graph grows.
READ
Retrieve entries by identity, type, grouping, or recency. Reads never alter state.
LINK
Record a typed relationship between two entries. Edges are first-class; the graph is the edges as much as the nodes.
REASON-ON
Return a coherent, model-shaped slice of the graph for a given scope: a briefing, not a search result. The operation that makes it memory the model thinks in, rather than a database it queries.
Mark an entry replaced by a newer one, and keep both, with the edge between them intact. This is the operation the whole standard turns on. An implementation that overwrites instead of superseding satisfies none of the axioms and conforms to nothing here. Keep the changed mind and the record can reason about its own history; overwrite it and you're left with storage.
The old decision stays in the graph.
supersedes edge. The model reading the graph months later sees not just the current decision but the decision it replaced and why, which is exactly the context that stops it from re-deriving, or re-making, a mistake already settled.What makes it more than storage.
Kept contradictions
A superseded decision isn’t an error to erase. The disagreement between old and new is signal. Real reasoning is born from contradiction, and the graph keeps it available rather than smoothing it away.
Provenance over content
Trust is derived from who wrote an entry, not from what the text claims. A poisoned entry can forge a citation; it cannot forge the identity that authored it. Trust travels on the author field.
Reasoning traces
The thinking that produced a decision MAY be captured as its own entry, linked to what it produced, so a later session inherits not just the conclusion but the road to it.
The primitive alone.
- The entry shape (§3)
- The five operations (§4)
- Append-only supersession (§2, §5)
- The epistemic guarantees (§6)
- Single model, single project
- What a model, an assistant, or an app embeds to remember
- Multi-agent coordination
- Spawning and orchestration
- Roles, hierarchy, routing
- Dashboards and visualization
- Hosting, tenancy, billing
- Everything a vendor competes and charges on
The standard is the ground. What you build to stand on it is yours to sell.
What “substrate-memory compatible” means.
The primitive is free so that everything can stand on it.
A memory the model reasons on should be infrastructure, the kind of layer every model, every assistant, and every long-running project can adopt without asking permission or paying a toll. Behind a license, it stays one company's product. Given away, it becomes a floor everyone can build on. The value was never in the schema, which anyone can copy in an afternoon. It's in what gets built on top, and that only grows when the floor is everywhere.
Free. Open. Universal.
The primitive above. Adopt it, embed it, re-implement it. A model can bake it into its weights; an assistant can ship it on a billion devices. No license, no toll, no permission. It has to win on merit, the way a standard does.
The cathedral. Where the work is.
Coordination, the fleet, the dashboard, the years of accreted lineage: a system with a time-moat rather than a schema. You can't copy it, because it isn't a spec. This is what a vendor builds, and what a vendor sells.
Latin for bond, link, that which binds. In mathematics, the bar over a repeating decimal — the mark that says these digits recur, indefinitely, as a unit. That's the architectural claim: a substrate that accumulates a project's intent across its whole life, holding parallel sessions together not by syncing them, but by giving them one authoritative surface to read from and write to. The sessions are ephemeral. The graph is not.
— the substrate that persists —