Skip to content

LlmWikis knowledge page

AI Dreaming Memory Management

AI Dreaming Memory is an offline, review-only consolidation pass over visible, source-backed project memory. It can propose duplicate merges, stale-claim cleanup, contradiction records, candidate summaries, and hot/cold memory deltas. It is proposal memory, not authority.

AI Dreaming Memory Management rule

Use dreaming when a wiki has enough reviewed material that humans need help seeing overlap, decay, contradiction, and promotion candidates. The useful output is not a confident answer; it is a queue of reviewable candidates with source pointers, redaction notes, and blocked-surface warnings.

Layer or surface Allowed use Promotion rule
raw/dream-runs/ Store raw dream-run artifacts, prompts, source set, hashes, date, model/tool notes, and run summary. Preserve for audit. Do not treat as reviewed wiki memory.
proposals/ai-dreaming/ Stage proposed summaries, duplicate merges, stale-claim updates, and hot/cold memory deltas outside canonical wiki indexing. Promote only after source trace, redaction, contradiction check, owner review, and lint.
wiki/contradictions/ Preserve conflicting claims surfaced during consolidation. Keep contradictions visible until a reviewer decides the public or internal claim boundary.
reports/dreaming-metrics/ Track review latency, stale candidates, rejected candidates, and false-promotion findings. Use metrics to improve process, not to auto-promote memory.
Reviewed wiki pages Receive only accepted deltas after review. No direct dream-run writes.
Hot memory: AGENTS.md, readme.human, and .uai Receive compact current facts only when the change affects routine pickup. A steward must approve and keep hot context concise.
Security, privacy, legal, release notes, production code, and archive deletion Protected surfaces. Dreaming may flag candidates; it must not modify them directly.

Review workflow

  1. Define the source set. Name the reviewed pages, session summaries, accepted source summaries, evidence logs, and named archive records that the pass may read.
  2. Record the raw run. Save run date, input paths, hashes or source pointers, tool notes, prompt scope, reviewer, and known exclusions under raw/dream-runs/.
  3. Stage candidates only. Put proposed summaries, duplicate merges, stale-claim cleanup, hot/cold deltas, and contradiction drafts in review queues.
  4. Check source and sensitivity. Every promoted claim needs a source link, redaction check, sensitivity label, contradiction state, and reviewer or owner.
  5. Promote narrowly. Write accepted deltas to reviewed wiki pages, hot Project Handoff memory, index/log records, or evidence reports only through the normal review path.
  6. Log the outcome. Record accepted, rejected, blocked, and deferred candidates, the checks run, and rollback or recovery notes.

Setup wizard fields

Field Why it matters
Mode Choose none, review-only candidate generation, scheduled review-only consolidation, or release-bound review.
Inputs Prevent broad memory grabs by naming only visible, source-backed inputs.
Run and candidate paths Keep raw artifacts separate from review queues and reviewed wiki pages.
Contradiction path Make conflicts durable instead of smoothing them away.
Review policy Require source links, redaction, contradiction preservation, lint, owner approval, and evidence updates before promotion.
Protected surfaces Block direct writes to reviewed wiki pages, hot handoff files, legal/security/privacy pages, release notes, production code, and archive deletion.
Metrics path Track whether the process is improving memory hygiene or adding noise.

Promotion checklist

  • The candidate names its source pages, source summaries, or archive records.
  • Private, legal, security, customer, credential, and personal data have been redacted or blocked.
  • Contradictions are preserved rather than flattened into a single confident claim.
  • The owner or reviewer approves the target surface and exact wording.
  • Lint, link, metadata, stale-claim, and discovery checks appropriate to the target surface have passed or are reported as blocked.
  • Index, log, evidence, and rollback notes are updated before the dream-run source leaves active review.

Current support boundary

LlmWikis teaches the review pattern. It does not currently ship a hosted AI Dreaming Memory service, autonomous repository writer, automatic LLM Wiki sync, automatic publication path, public MCP or write API, SDK, CLI, certification, endorsement, hidden model memory, model-weight training, or UAI-1 conformance evidence. UAIX remains canonical for AI Memory and Project Handoff definitions.