WikiMemory

Open the deck with your friendly neighborhood agent

Open this presentation in your browser:
https://la3d.github.io/nuggets/projects/sai/updates/2026-08-31/
Before helping me, inspect the page's raw HTML. Find the embedded

W3C Web Annotation JSON-LD and load its per-slide learning context.
I will navigate the deck myself and may skip around. When I ask a question,

use the slide number or title I provide to select the matching annotation.

Respond conversationally and only to what I ask. Draw on the core claim,

caveats, and sources when useful. Do not lecture through the deck, advance,

or quiz me automatically. If I do not identify the slide, ask which one I am

viewing.

A memory complaint became an agentic design pattern

The surprise was not that a model could remember—it was that an agent could maintain external memory.

Karpathy on distracting memory · Elvis Saravia on methodical file-based memory · Karpathy, “LLM Knowledge Bases” · the “LLM Wiki” idea file.

Compile knowledge once; improve the artifact over time

Andrej Karpathy, “LLM Wiki” · Harrison Chase, “Wiki Memory”.

Every model call begins now; science began long ago

Research advances by carrying evidence, failed attempts, critique, decisions, and learned procedures across people and time.

WikiMemory is a convergence—not a sudden invention

Our synthesis: three historical lineages converge in a harness-operable knowledge substrate.

Bush imagined trails—not a better filing cabinet

Vannevar Bush, “As We May Think”, 1945.

Hypertext was designed as a cognitive environment

Doug Engelbart Institute, NLS / Augment · Project Xanadu history.

The Web began as a continuity problem at CERN

People rotated through. Projects evolved. Information existed—but could not be found.

Turnoverexpertise and project history left with short-term staff

Entanglementreal work crossed the formal organizational hierarchy

Changebooks and trees became stale as dependencies shifted

The proposal was for evolving organizational memory—not merely online publishing.

Tim Berners-Lee, “Information Management: A Proposal”, March 1989 / May 1990 · original W3C diagram.

Berners-Lee proposed typed nodes and links

Before the Web had pages at planetary scale, its proposal described a flexible, distributed knowledge graph.

Berners-Lee, “Information Management: A Proposal,” 1989 · sections: Linked information systems · CERN Requirements · Data analysis.

A typed link says why two things are connected

Dan Brickley, “Nodes and Arcs 1989–1999: The WWW Proposal and RDF: Then and Now” · informal W3C discussion, 1999.

The Web began as a read–write medium

Wikis did not invent read–write hypertext; they recovered an ambition the mainstream Web largely left behind.

1989 client/server proposal · W3C, “WorldWideWeb: the first web client” · Berners-Lee, “Realising the Full Potential of the Web”.

Wiki made knowledge maintenance a social process

Ward Cunningham, “Design Principles of Wiki”, WikiSym 2006.

Links became infrastructure for collective memory

MediaWiki history · Wikipedia launched in January 2001; double-bracket “free links” emerged in the early Wikipedia toolchain.

Plain text became both prose and data

YAML specification history · Gruber, “Introducing Markdown” · Preston-Werner, “Blogging Like a Hacker”.

The wiki became a personal knowledge graph

The Story of TiddlyWiki · Roam Research · Obsidian · Balog & Kenter, 2019.

Luhmann built a network of arguments—not a filing system

Johannes F. K. Schmidt, “Niklas Luhmann’s Card Index” · “The Fabrication of Serendipity”.

A digital PKG preserves the mechanics—not the furniture

The graph view displays a network; the linking practice creates one worth traversing.

Niklas Luhmann Archive · Obsidian internal links.

Obsidian makes the substrate inspectable

The vault supplies the medium; linking and maintenance practices turn it into memory.

Obsidian internal links · Obsidian properties.

The new actor is not another reader—it is a maintainer

WikiMemory is not a new storage idea. It is a new actor in an old medium.

The shift became widely visible through Karpathy’s “LLM Wiki”, April 2026.

RAG retrieves fragments; WikiMemory preserves synthesis

Not a replacement: raw-source retrieval remains essential for evidence, verification, and details lost in compilation.

WiCER finds blind compilation can discard critical facts; one or two evaluate–refine iterations recover much of the lost quality.

Structure turns a folder into an operable knowledge space

Design synthesis. Plain links remain the navigation baseline; add typed predicates when the relation changes query, validation, explanation, or workflow.

index.md turns hierarchy into a context budget

# Scientific WikiMemory

Claims

Synthesized claims, evidence, and status.
Methods

Procedures, assumptions, and validation.
Experiments

Runs, outcomes, failures, and decisions.
Sources

Papers, data, code, and conversations.

Context grows with relevance—not with the size of the corpus.

Google Cloud, Open Knowledge Format v0.2, §8 “Index files”.

A WikiMemory document is a typed node—not an isolated page

Design synthesis: typed note roles and edges make the inquiry path inspectable; the concept note coordinates the neighborhood rather than absorbing every assertion.

A claim note keeps maturity separate from epistemic status

---
type: claim
maturity: draft
epistemic_status: contested
creator: researcher-or-agent-id
context: inquiry/compilation-gap
supported_by: ["[[paper-17]]"]
contradicted_by: ["[[run-042]]"]
review_after: 2026-10-01
---

Maturity asks how developed the note is. Epistemic status asks how well the claim is currently supported.

Illustrative Obsidian-compatible claim note—not a canonical schema for llm-wiki-colab or every vault.

The neighborhood is the transition space; the realized sequence is the trail

Markov-chain connection: notes can be modeled as states and typed links as possible transitions. This is a design analogy—not a claim that the current plugin implements one particular Markov algorithm.

The harness turns a writable vault into a memory system

Persistence is an externalized capability; reliability comes from how the harness governs it.

Zhou et al., “Externalization in LLM Agents” · “The Living Wiki” treats the vault schema as a procedural Skill.

A WikiMemory write changes a neighborhood—not just one file

LA3D-LLM-Agents, llm-wiki-colab · Saboia Moreira et al., “Beyond Memory”.

Readable predicates compile into an operable graph

supports: "[[Claim-X]]"
dependsOn: "[[Dataset-Y]]"

[Claim X](Claim-X)
([*supports*](Edge-Types#supports))

Edge-Types.md vocabulary · KG build pipeline · forward predicates are authored; inverses are materialized.

Skills make scientific maintenance explicit

Skills in llm-wiki-colab · each write procedure delegates to a shared Verification Gate.

A federation keeps memory local—and expertise reachable

Share an honest capability description; consult the source wiki; avoid copying every project into private context.

LA3D LLM Agents organization profile · federation index.

One knowledge space, many artifacts

Stable IRIs connect what each owner keeps authoritative; projections and catalogs carry links—not duplicate copies of every artifact.

Live now: this presentation’s AnnotationPage. Prospective: deployed WikiMemory manifests and a cross-project discovery catalog.

Web Annotation · PROV-O · SKOS · CiTO · DCMI Terms.

Follow the link; load the neighborhood

JSON-LD alone is not FAIR. Private knowledge stays private; shared semantics, access policy, provenance, licensing, and governance determine what can travel.

Wilkinson et al., FAIR Guiding Principles · W3C Data on the Web Best Practices · standards: Web Annotation · PROV-O · SKOS · CiTO · DCMI Terms.

Recent work is filling in different parts of the system

Chase, 2026 · WiCER · LLM-Wiki · WikiLoop · TrajWiki · Beyond Memory · WikiSkill.

An agent-maintained wiki can compound error, too

“Memory as Metabolism” highlights entrenchment and contradictory evidence · TrajWiki uses immutable snapshots plus ADD / REVISE / DEPRECATE operations.

WikiSkill compiles experience into validated procedure

Key result: persistent wiki knowledge materially improves skill evolution; accepted skills can transfer across models.

Tang et al., “WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution”, Google Research, Aug. 2026.

Scientific memory must preserve why we believe—or doubt

A scientific second brain should not only remember conclusions; it should preserve the routes by which conclusions remain contestable.

What would make WikiMemory scientifically trustworthy?

Can an inspectable wiki become the shared memory substrate through which scientists and agents learn together?