Deep Cognitive Archive

The Deep Cognitive Archive is AIWikis.org's local name for the slower memory layer behind public pages and compact handoff packets.

It keeps durable reasoning where future humans and AI agents can inspect it without loading every raw file or turning the public site into a transcript dump. The model is applied here as AIWikis dogfood. It is not a UAIX certification claim, a public LLMWikis baseline requirement, or a hosted validation service.

The Three Layers

Layer	Job	What belongs there
Public site	Polished claim surface	Reviewed summaries, routes, package ownership notes, source boundaries, and current guidance.
AI Memory or Project Handoff packet	Compact active packet	`AGENTS.md`, `readme.human`, and selected `.uai` files that are useful at chat start or handoff time.
Deep archive	Slower reasoning store	Raw archive files, source proxies, compiled wiki nodes, graph edges, decision history, rejected alternatives, contradictions, and modelling limits.

The public site answers "what should a reader know now?" The compact packet answers "what does the next agent need loaded immediately?" The deep archive answers "why is the system shaped this way, what changed, and what remains uncertain?"

What Goes Where

Use content/pages/ for reviewed public Markdown. Use AGENTS.md, readme.human, and .uai/ for operational handoff. Use raw/system-archives/, wiki/, and config/graph.jsonld for the durable memory layer.

For archive memory, keep site and bucket boundaries visible:

Source	Long-term archive pattern
UAIX.org	`raw/system-archives/uaix/agent-file-handoff/Archive/YYYY-MM-DD/Content-or-Improvement/...`
LlmWikis.org	`raw/system-archives/llmwikis/agent-file-handoff/Archive/YYYY-MM-DD/Content-or-Improvement/...`
AIWikis.org	`raw/system-archives/aiwikis/agent-file-handoff/Content-or-Improvement/YYYY-MM-DD/...` plus the local archived intake file.

Content means candidate editorial or public material. Improvement means audits, QA reports, strategy, bug notes, architecture suggestions, and other work-to-apply. Archive means the file has already been dispositioned and is dead to routine intake unless a human explicitly names it.

Deep-Memory Node Shape

AIWikis source-side wiki nodes use local metadata for identity, lineage, confidence, linting, and routing. That shape is an AIWikis extension, not a universal LLMWikis or UAIX mandate.

---
uai_id: "uuid"
type: "concept"
lineage:
  - "raw/source-or-archive-record.md"
confidence: 0.84
last_linted: "2026-04-29T01:39:06Z"
aliases:
  - "alternate name"
---

Numeric confidence is local wiki metadata. Public guidance should avoid pretending that every project has a universal mathematical confidence score unless the project has an explicit scoring policy. For public copy, qualitative labels such as reviewed, source-linked, source-needed, contradicted, stale, and archived are often safer.

Ingest Loop

Preserve the original source in raw/ or raw/system-archives/ without editing it.
Create or update a source-proxy page in wiki/sources/.
Update the concise concept page only with stable conclusions.
Create or extend a deep-memory concept or decision shard when the "why" would overload the public page.
Update wiki/index.md, config/graph.jsonld, and the relevant log node.
Promote only compact accepted truth into .uai files when the active handoff packet actually changes.

This keeps the archive useful without making every thought public, authoritative, or active by default.

Project Handoff Export Loop

AIWikis now treats the handoff packet as a buildable artifact, not only a set of hand-maintained files. tools/generate-handoff-export.ps1 assembles a local dist/handoff/aiwikis-project-handoff/ packet from root handoff files, .uai files, recommendation docs, active intake scaffolding, and selected reviewed wiki nodes marked for export.

The export is local dogfood evidence. It does not create UAIX certification, hosted generator support, endorsement, or UAI-1 conformance evidence. Its job is narrower: prove that reviewed wiki memory can feed a compact Project Handoff packet without stuffing raw archives into chat-start context.

Retrieval Loop

Agents should read wiki/index.md first, choose the smallest useful page set, and go deeper only when the task needs provenance, tradeoffs, contradictions, or historical decisions. The deep archive is a routing layer, not a context-window stuffing strategy.

Boundaries

AIWikis archive copies are memory and implementation evidence; they are not canonical UAIX or LlmWikis truth.
Raw reports are not public content until rewritten as reviewed summaries.
Graph export is an interoperability enhancement layer, not a promise that every project exposes RDF or JSON-LD.
The archive should preserve contradiction and rejected alternatives, but it should not become a dump of every AI-generated sentence.
SDKs, CLIs, hosted validators, certification, and endorsement remain unsupported unless the source sites publish that evidence.

Next Useful Routes

Start Here A task-first reading path for AIWikis.org, separating newcomer learning, source-memory lookup, maintainer workflow, and AI-agent retrieval.
Topic Index A tag-oriented index for LLM Wiki, AI memory, UAI, source governance, crawling, and retrieval topics.
Source Map AIWikis source-governed page for durable AI memory, evidence routing, and agent-readable retrieval.