LLM Wiki vs. UAIX Project Handoff

LLM Wiki vs. UAIX Project Handoff: Two Ways to Keep AI Work From Losing Its Memory

LLM Wiki helps knowledge accumulate. UAIX Project Handoff helps project state travel. They are related, but they should not be treated as the same layer.

Reader boundary

This report compares two durable-context patterns. It does not create a new UAIX support claim, validator, generator, SDK, CLI, certification path, or endorsement program. Use Project Handoff for the current practical handoff pattern, and use UAI-1 plus the Validator when a local handoff becomes public exchange evidence.

The Shared Problem

AI-assisted work often loses the state it just built. A model may spend an hour reading code, notes, policies, and architecture decisions, then that working context disappears when a chat ends, a context window compacts, or a different agent takes over. Retrieval-augmented generation addresses part of the problem by retrieving relevant source chunks at query time, but the original RAG framing still keeps the retrieved material separate from the model and makes provenance and updating explicit design concerns.⁵

Karpathy’s LLM Wiki pattern attacks the continuity problem by asking the LLM to build and maintain a persistent Markdown wiki between raw sources and future questions. UAIX Project Handoff attacks a different continuity problem: how the next AI, team, vendor, or model should safely continue work in a real repository. Both move important context out of disposable chat. They differ in what that context is allowed to govern.^{1 6}

What LLM Wiki Optimizes

LLM Wiki is a knowledge-compounding pattern. Karpathy describes three layers: immutable raw sources, an LLM-owned wiki of generated Markdown pages, and a schema file such as AGENTS.md or CLAUDE.md that tells the LLM how to ingest, query, and maintain the wiki. The important shift is that the wiki becomes a durable, edited artifact. Summaries, entity pages, concept pages, comparison pages, indexes, and logs keep improving instead of being regenerated from scratch every time someone asks a question.¹

That makes LLM Wiki strongest when the work is open-ended research, long-form reading, competitive analysis, internal knowledge management, or repeated synthesis across many documents. Its center of gravity is learning and sense-making. The wiki can inform engineering work, but it should not automatically become engineering instruction.

What Project Handoff Optimizes

UAIX Project Handoff is a repository-continuity pattern. It starts with a root AGENTS.md file and a local .uai/ folder that contains typed context records such as context.uai, stack.uai, constraints.uai, architecture.uai, decisions.uai, progress.uai, operations.uai, and style.uai. The root file lists explicit @uai[] references so the next AI knows what to load before broad edits.^{6 7}

The handoff pattern is intentionally more operational than encyclopedic. It asks the next AI to summarize the project, confirm which files loaded, name hard constraints, identify intended touchpoints, and state verification checks before making broad changes. That first response makes stale assumptions, unsupported claims, missing files, or risky edit paths visible early.⁶

Side By Side

The Clean Boundary

The safest architecture is not to choose one pattern and discard the other. Use LLM Wiki as the epistemic layer: it is where research, source notes, contradictions, comparisons, and domain background can grow. Use UAIX Project Handoff as the execution layer: it is where project purpose, stack, current state, operating commands, and constraints become the brief the next AI must load before acting.

When a claim leaves the local repository and becomes release-facing or interoperability-facing, move into the transmission and evidence layer: UAI-1, schemas, registry records, examples, validator results, conformance packets, changelog entries, and implementation records. UAIX’s Standards Fit guidance uses the same boundary logic for adjacent protocols: UAI-1 records the portable exchange; other systems keep their own runtime jobs.⁹

Governance Risks

The LLM Wiki risk is epistemic drift. A clean generated page can look more authoritative than the source it summarizes. Teams should keep raw sources immutable, cite them, review important synthesis, and treat the wiki as a working interpretation, not unquestioned ground truth.

The Project Handoff risk is operational drift. A stale constraints.uai, missing operations.uai, or outdated progress.uai can mislead the next agent. That is why UAIX keeps hard constraints separate, requires explicit context loading, and frames linked .uai files as context rather than authority that can override the human, system instructions, policy, or repository rules.⁶

Takeaway

LLM Wiki and UAIX Project Handoff are complementary memory systems. The wiki preserves what the team is learning. The handoff preserves what the next actor needs to do safely. A mature AI engineering workflow uses both, but keeps the wiki informative, the handoff governing, and public support claims tied to UAI-1 evidence.

Works Cited

1. Andrej Karpathy, LLM Wiki, GitHub Gist, created April 4, 2026.
2. AGENTS.md, A simple, open format for guiding coding agents.
3. OpenAI Developers, Custom instructions with AGENTS.md – Codex.
4. OpenAI Developers, Customization – Codex.
5. Patrick Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, arXiv:2005.11401.
6. UAIX, Project Handoff.
7. UAIX, AGENTS.md .uai Linking Specification.
8. UAIX, UAI-1.
9. UAIX, Standards Fit.
10. UAIX, Validator.