Tony Seale's Post

Tony Seale · article · 2026-07-05 source ↗ Reader 🖥️

Agent Memory Agent-Native Design Agent-to-Agent Protocols Context Engineering Knowledge Base Neurosymbolic AI Organizational Legibility

Durable agent memory should fuse the two kinds of human memory — a markdown wiki for the episodic (what happened) and a formal ontology for the semantic (what things mean) — in one layered file format, Vault-LD (markdown plus optional YAML-LD frontmatter), so that any note can progressively acquire structure and meaning and a private vault can scale, unchanged in substrate, from a personal second brain up to an open, decentralised agentic web.

practical · argumentative (manifesto), with expository passages that define Vault-LD — A LinkedIn manifesto: promotional, first-person, 'you heard it here first', announcing an open standard the author's own company already runs on. Practical because it advocates an action and a design — build agent memory this way — rather than establishing a truth for its own sake. Persuasion is carried by analogy, a psychology borrow, and n=1 practitioner testimony rather than by evidence.

Structure

Part	Role
Two-trends hook: agent builders converge on markdown wikis for memory; enterprises converge on semantic layers/ontologies to ground agents — 'nobody has put them together'	Frames an unexploited opportunity and the author's claim to priority
The synthesis claim: a markdown wiki can be combined with a formal knowledge graph and ontology	Central thesis assertion
Episodic vs semantic memory mapping (Tulving borrow): markdown = episodic substrate, ontology = semantic meaning	Theoretical grounding by analogy to cognitive psychology
Testimony + reveal: 'we've run our company on this for a year... now publishing it as an open standard' — Vault-LD	Evidence-by-practitioner-experience plus the product announcement
Mechanism definition: markdown + YAML-LD frontmatter = prose (the prompt) + structure (typed queryable fields) + semantics (linked data to a formal ontology), each layer optional until needed	Operationalizes the thesis into a concrete file format
JSON-LD-on-HTML analogy: the web became a global knowledge graph via JSON-LD islands; Vault-LD applies the same move to agent memory	Legitimation by historical precedent
Progressive-enrichment + scaling telos: prose -> typed fields -> linked meaning -> graph projection (inference/validation/query/integration) -> DPROD data products -> second brain -> team -> enterprise -> agentic web	Displays the pattern's range and ambition
Chiasmus closer: 'Ontology is how agent memory grows up. Agent memory is how ontology goes everywhere.'	Rhetorical crystallization of mutual reinforcement

Key terms

Vault-LD
The proposed open standard: markdown files with optional YAML-LD frontmatter, holding prose, structure and semantics in one file, so a vault of notes can project into a formal knowledge graph.
Episodic vs semantic memory
A design split borrowed from psychology: episodic = what happened (meetings, decisions, plans, notes) handled by markdown; semantic = what things mean (identity, type, relationship, constraint) handled by ontology.
YAML-LD frontmatter
The structure-plus-meaning layer placed atop a markdown file: YAML supplies typed fields a machine can query; LD (linked data) connects those fields back to a formal ontology.
Ontology Organizational Legibility
The semantic layer supplying identity, type, relationship, constraint and meaning; 'how agent memory grows up' — the shared formal vocabulary a vault's fields link into.
Context graph / semantic layer
What enterprises (Databricks, Snowflake, Google, Microsoft, Amazon) are building to ground their agents; the graph a vault projects into where 'inference, validation, query and integration are already waiting'.
DPROD data products
The packaging by which a vault is exposed for integration, letting the pattern scale out from team memory to enterprise integration to an agentic web.

Propositions

explicitThe simplest way to give an agent durable memory is a markdown wiki.
explicitMajor enterprise vendors (Databricks, Snowflake, Google, Microsoft, Amazon) are converging on semantic layers, context graphs and ontologies to ground their agents.
explicitA markdown wiki and a formal knowledge graph/ontology do not need to be separate ideas — they can be combined.
extendsLLM Knowledge Bases · p1 — Karpathy holds that LLMs compile raw documents into structured, interlinked markdown knowledge bases. Vault-LD takes that markdown-KB and extends it with a formal semantic layer: YAML-LD frontmatter projecting the same files into a knowledge graph/ontology, so the base gains inference and validation the flat markdown version does not attempt.
explicitcontestedMarkdown is a good substrate for episodic memory (events, decisions, notes); an ontology supplies semantic memory (identity, type, relationship, constraint, meaning).
supportsWow, this tweet went very viral! · p2 — Karpathy: LLMs are better used to manipulate knowledge (stored as markdown) than to manipulate code. Seale's claim that markdown is 'a beautiful substrate' for episodic memory and 'the prompt the model reads' supports the same premise — markdown as the right substrate for LLM-manipulated knowledge.
contradicts ←Paweł Huryn's Post · p4 — Direct disagreement on which memory type markdown serves: Huryn maps markdown to semantic memory and a CSV to episodic memory, while vault-ld maps markdown to episodic memory (events/decisions/notes) and a separate ontology layer to semantic memory (identity/type/relationship). Same building blocks, opposite assignment.
explicitVault-LD = markdown plus optional YAML-LD frontmatter, carrying prose (what the model reads), structure (typed queryable fields) and semantics (linked data to a formal ontology) in one file, each layer optional until needed.
explicitBecause the web became a global knowledge graph by adding JSON-LD islands to HTML, the same annotation move (YAML-LD frontmatter on markdown) will work for agent memory.
explicitdevelopedA vault can progressively enrich (prose -> typed fields -> linked meaning -> graph projection) and, exposed as DPROD data products, scale unchanged from a private second brain to team memory to enterprise integration to an open decentralised agentic web.
extends ←Presentation: Autonomous Data Products for the Autonomous Era: Rethinking Data Architecture for GenAI · p4 — extract-vault-ld gestures at personal vaults scaling into 'DPROD data products' from private to enterprise without developing the mechanism; this talk supplies the concrete lifecycle/kernel/policy architecture that vision leaves undeveloped.
implicitCombining the two layers is net-positive — the value of adding structure and semantics exceeds the ongoing cost of authoring them and keeping them in sync with evolving prose.
implicitSomeone with knowledge-engineering skill defines and maintains the shared formal ontology, and it stays coherent as many notes link into it.
implicitThe web's structured-data adoption was a success worth emulating, and that adoption will transfer to private and team vaults.
implicitdevelopedThe inference/validation/query/integration payoff is real and worth the ontology overhead.
refines ←Presentation: From Retrieval to Reasoning: Building Production-Ready Agentic AI Systems with Knowledge Graphs · p12 — Vault-LD asserts the graph/ontology overhead is generally worth its payoff; Shum adds the conditionality that it is only worth it when a specific, demonstrated benefit (org-specific context, efficiency at scale) is present.

Arguments

You should architect agent memory as a fusion of a markdown wiki and a formal ontology, via Vault-LD (markdown + optional YAML-LD frontmatter). strength: medium

Explicit premises:

Agent builders are converging on markdown wikis for durable agent memory.
Enterprises are converging on semantic layers and ontologies to ground their agents.
Episodic memory (markdown) and semantic memory (ontology) are complementary kinds, mirroring how human memory is organized.
The web already succeeded by adding JSON-LD islands to HTML; Vault-LD applies the same incremental move to markdown.
We have run our own company on this exact combination for over a year and it is an extremely powerful pattern.

Implicit premises:

The maintenance and sync cost of the structure and semantic layers is low enough to be dominated by their benefits ('each layer optional until you need it').
A coherent shared ontology already exists and is kept coherent by someone with the skill to do it.
Adoption of the semantic layer will happen even though it is optional and lacks a search-ranking-style forcing function.
Partial, optional semantic coverage still yields reliable inference over the whole vault.
Mapping machine memory onto the human episodic/semantic distinction is a valid architectural guide, not merely a suggestive metaphor.
The quality of the underlying notes is high enough that projecting them into a formal graph confers useful, not falsely-authoritative, structure.

Inferences:

If maintaining the ontology is the hard part, the bottleneck moves from note-taking convenience to scarce knowledge-engineering labour — the opposite of 'the simplest thing that could possibly work'.
'Optional until you need it' plus ordinary human laziness predicts most vaults stay mostly prose, leaving the graph sparse exactly where semantics were supposed to add value.
A one-year, single-company track record published by the standard's own author is n=1 and self-interested evidence.
Hallucinated or drifting notes, once given typed fields and ontology links, inherit machine-queryable false authority — the semantic layer can propagate bad facts as readily as it validates good ones.

Narratives

The web became a distributed global knowledge graph by adding islands of JSON-LD to HTML pages; Vault-LD applies the same move to agent memory by putting YAML-LD frontmatter at the top of markdown files.
→ Legitimation by historical precedent — the key rhetorical move of the piece, framing incremental structured annotation as a proven, inevitable path.
Psychologists distinguish episodic memory (what happened) from semantic memory (what things mean); markdown is the substrate for the first, ontology for the second.
→ Theoretical grounding — borrows a respected cognitive-science taxonomy to make a technical design choice feel principled rather than arbitrary.
'We've spent just over a year running our company on exactly this combination. It is an extremely powerful pattern.'
→ Evidence-by-practitioner-experience — supplies a track record where data is absent.

Questions addressed

✅ How do you give an agent durable memory?
A markdown wiki, optionally enriched with typed fields and an ontology.
✅ Do the 'markdown wiki' and 'enterprise ontology' approaches have to be separate?
No — they are the episodic and semantic halves of one system and can be combined.
✅ How, concretely, do you carry prose, structure and meaning in one artifact?
Markdown for prose, YAML for typed structure, LD for linked-data meaning — YAML-LD frontmatter on markdown files.
⭕ Who defines and maintains the shared formal ontology, and at what cost?
Never raised; the ontology is treated as a substrate that 'supplies' meaning as if it were free-standing and free.
⭕ What does the inference/validation/query/integration payoff actually deliver — a worked example?
Asserted as 'already waiting' with no example of a question answered or an error caught that flat markdown could not manage.
⭕ What forces adoption of the semantic layers, given they are optional and the vault works without them?
Unaddressed; the web analogy assumes an adoption dynamic without naming the vault-world equivalent of search-ranking pressure.
⭕ How do typed fields and ontology links stay in sync as the prose they describe evolves?
Drift is invisible in the piece; 'each layer optional until you need it' quietly assumes the layers, once added, stay correct.

Gaps

incompleteOntology maintenance and curation cost is entirely unaddressed. The post treats the ontology as a layer that 'supplies' semantic memory, but says nothing about who authors the shared formal vocabulary, how it is governed as many notes link into it, or that this is scarce knowledge-engineering labour. If maintaining the ontology is the hard part, the bottleneck moves from markdown convenience to specialist curation — undercutting the linked 'simplest thing that could possibly work' framing.
Read next: Read the historical adoption record of formal ontologies and the Semantic Web (OWL/RDF, schema.org governance) for why ontology upkeep, not authoring notes, is the recurring failure point. Compare with this repo's own knowledge/ layer, which runs the same episodic+semantic split but deliberately diverges from Vault-LD: it keeps the ontology in an external, human-curated registry (concepts.yaml) rather than inline YAML-LD, and uses positional refs (slug#p3) rather than linked-data URIs. That a working instance of the pattern chose an external registry is direct evidence that inline 'one file, each layer optional' is not the only viable architecture — and that someone still has to own the semantic layer by hand.
incompleteThe inference payoff is asserted but never demonstrated. 'Project into a graph, where inference, validation, query and integration are already waiting' names four benefits without a single worked example of what an agent can now answer, or what a validation constraint would catch, that a plain markdown wiki could not. The most consequential claim in the piece — that the ontology overhead buys real capability — is left as a promise.
Read next: Seek the linked Vault-LD spec for a concrete example, then pressure-test it against the counter-claim in the corpus (Karpathy: an LLM-maintained markdown index is sufficient for complex Q&A at ~100-article scale without any graph). Ask specifically: what query, inference or validation justifies the ontology, and at what vault size does flat markdown actually break?
incompleteThe net-positive assumption skips the sync/drift ledger. Keeping typed fields and ontology links in agreement with prose as the prose evolves is ongoing cost, and 'each layer optional until you need it' in practice tends to mean 'skipped' — which leaves the graph sparse exactly where semantics were supposed to add value, and leaves the annotated minority prone to silent divergence from the text it describes.
Read next: Read on documentation drift / doc-rot and on the divergence between described and enforced structure over time (cf. the static-sensors debate on markdown guides vs enforcement rules). Consider whether YAML-LD frontmatter has the same rot dynamics as any duplicated metadata, and what would keep it honest.
incompleteThe JSON-LD-on-HTML analogy imports an inevitability it has not earned. The web's structured-data adoption was driven by a concrete incentive — search engines reward JSON-LD with ranking — and is itself uneven and gamed. The post identifies no equivalent forcing function for private or team vaults, yet treats 'the same move' as a settled outcome rather than a hopeful one.
Read next: Examine schema.org adoption data and its SEO dynamics — how much of the web's 'knowledge graph' is genuine semantics versus ranking-driven boilerplate — and ask what a vault-world analogue of search-ranking pressure would even be. If none exists, the adoption story needs a different engine than the analogy provides.
incompleteNo treatment of the trustworthiness of the memory itself. Agent memory is largely LLM-written; a hallucinated or drifting note, once given typed fields and ontology links, inherits machine-queryable false authority. The semantic layer is presented purely as validation upside, with no acknowledgement that it propagates bad facts as readily as it structures good ones, or that memory poisoning becomes more consequential once memory is formally queried.
Read next: Read on agent-memory poisoning and knowledge-base trust (how errors propagate through an LLM-maintained base over time). Distinguish validation that catches contradictions from structure that merely launders unverified claims into confident graph facts.

Closes gaps elsewhere

Three more static code analysis sensors · g8 — That gap warns that migrating meaning into machine-only enforcement tooling (dependency-cruiser rules replacing a markdown structure guide) destroys the human-readable prose that teaches — the 'tests are the spec' failure. Vault-LD's core move directly answers it: keep prose AND typed structure/semantics in one file so the two epistemic jobs — teaching (markdown) and constraining/validating (YAML-LD) — coexist per-file rather than one substituting for the other.
LLM Knowledge Bases · g4 — That gap notes Karpathy assumes markdown is the right format and never considers whether richer representations (knowledge graphs, ontologies) would enable capabilities flat markdown cannot. Vault-LD is precisely that consideration and a proposed answer: keep the flat markdown but add optional YAML-LD frontmatter so the same files project into a formal knowledge graph/ontology, unlocking inference/validation/query without abandoning the prose.
LLM Knowledge Bases · g5 — Partially closes it. That gap flags that Karpathy's workflow is solo-only, with no collaborative or multi-agent perspective. Vault-LD supplies exactly that multi-scale story — team memory, enterprise integration and multi-agent exposure via DPROD data products up to an agentic web — though it still omits the version-control/provenance mechanics the gap also asks for.