Persistent knowledge that compounds — not RAG that re-derives.
microsoft/amplifier-bundle-llm-wiki · default-on in amplifier-foundation · May 2026
Every RAG query starts from zero — chunk, embed, retrieve, reason, throw it away. The model never accumulates structured understanding of your sources. Andrej Karpathy's LLM Wiki pattern proposes the inverse: the wiki is the compiled artifact. The LLM owns the bookkeeping. Humans curate.
Sources stay raw. Every query re-runs the same expensive understanding work. There is no place where knowledge accumulates.
Raw inputs are mined incrementally into entities. The wiki cross-references itself. Queries read from the compiled layer.
The model maintains schema, links, freshness. People decide what's worth ingesting and review what landed.
A composable workflow surface: each /wiki-* mode is a focused step. They share orientation but never load it unless one is active.
Contributed only by /wiki-init's contributes.agents. Mounts only when that mode is active. Body loads only when delegate is called against it. The actual scaffolder behind /wiki-init.
The bundle is a working, ecosystem-integrated implementation of the gist — wired into how Amplifier already handles modes, behaviors, and bundle composition.
A normal cadence: scaffold once, then ingest as sources arrive, lint before shipping, publish on demand, query whenever. No mode is required — each is independently useful.
Linear by default, but the modes are independent: you can lint without publishing, query without ingesting, re-run init to evolve the schema. The cycle is a guide, not a state machine.
The bundle's persistent footprint is zero because YAML config never enters LLM context — it's processed at runtime to wire up structural discovery, then discarded. Only mode bodies and contributes.context files reach the model, and both are mode-gated. See amplifier-bundle-modes/context/mode-schema-reference.md §9.7 for the mechanism; this bundle is its reference implementation.
| Session state | Tokens in LLM context |
|---|---|
| Bundle composed; no wiki mode active | 0 |
| Any one wiki mode active | mode body + ~500 shared orientation |
| /wiki-init active and delegating to the agent | + agent body (only during the delegate call) |
Adoption tax for users who never touch the wiki: nothing.
The bundle is default-on in amplifier-foundation. Any standalone that composes foundation — amplifier-dev, with-anthropic, with-openai, and friends — picks it up transitively. The five /wiki-* modes are already discoverable via /modes. Nothing to install.
If you're on a different base and want the wiki bundle as an overlay (works directly with amplifier-app-cli, or on top of any other bundle that has mode infrastructure):
Either way, bootstrap a new project with a brief:
The init flow delegates to wiki-policy-designer, which interviews you about schema, publish target, and viewer — then scaffolds .wiki/, raw/, the publish script, and the project's AGENTS.md entry.
One plausible shape — design-review transcripts in, an ADR-style decision wiki out. The same flow applies to research-paper summaries, support runbooks, weekly meeting notes, or any compiled-artifact workflow.
Karpathy proposed the shape; multiple implementations have explored corners of it. The Amplifier bundle is informed by that lineage — the contribution is making it composable, ecosystem-native, and dormant-by-default.
None of these are direct dependencies. They are reference points — the Amplifier bundle's distinctive contribution is the zero-cost-when-dormant integration, not the wiki workflow itself.
If you see /wiki-init, /wiki-ingest, /wiki-lint, /wiki-publish, /wiki-query — you already have it. Otherwise, overlay it:
amplifier bundle add --app git+https://github.com/microsoft/amplifier-bundle-llm-wiki@main#subdirectory=behaviors/llm-wiki.yaml
/wiki-init <describe what you want to track and ship>
github.com/microsoft/amplifier-bundle-llm-wiki
amplifier-bundle-modes/context/mode-schema-reference.md §9.7
amplifier/docs/MODULES.md (llm-wiki entry)
If you build a wiki with it, the team would love to see the schema you converge on. The patterns aren't fixed — /wiki-init is designed for any compiled-artifact workflow.