Skip to content

LlmWikis knowledge page

Multi-Agent Transactive Memory

The LLMWikis governed MATM profile is a practical implementation profile for teams that want agent-generated experience sharing without direct worker writes. It is not a universal standard, conformance program, or certification.

Public/private bridge boundary

MATM trajectories, private retrieval records, tool outputs, hidden reasoning, customer data, credentials, and workspace-specific traces are unsafe as public pages. Public LLMWikis guidance may contain only sanitized procedures, schemas, templates, synthetic examples, trust metadata, and support-boundary language after promotion gates pass.

Bridge Requirements

  • Decouple private runtime evidence from public handbook guidance before drafting public copy.
  • Rewrite private lessons as general patterns with source-traced metadata, redaction status, lane review, and explicit non-claims.
  • Use state-conditioned and scope-filtered retrieval; retrieval visibility is never authorization, purchase authority, write permission, or public evidence by itself.
  • Reject direct worker durable writes. Workers submit proposals; curators and reviewers decide durable records and public promotions.
  • Publish synthetic or redacted trajectory summaries only when they teach a pattern and cannot expose private trace detail.

Base Wiki Contract

  • MUST inherit the base LLM Wiki contract before enabling MATM: Markdown concept files require type frontmatter, and Memory Events or Memory Records are not replacements for reviewed concept pages.
  • MUST keep wiki/index.md and wiki/log.md reserved; multi-project root files name project-level index/log counterparts instead of listing every file or every event. Child indexes and logs carry scoped detail.
  • MUST preserve unknown frontmatter keys and treat broken local links as draft warnings that become strict release failures.
  • MUST partition MATM scope, authorization, retrieval, and audit context by client, workspace, and project or site namespace.
  • MUST preserve split-memory boundaries: MATM durable records do not replace local .uai / AI Memory startup memory, and consumer agents read hot local memory first for current truth, constraints, blockers, and recovery posture.

Profile Rules

  • MUST keep worker agents proposal-only for durable memory. Workers emit Memory Events; they do not write durable records directly.
  • MUST keep hidden chain-of-thought and private reasoning out of durable-memory requirements.
  • MUST map core or local scope names into the six-scope practical profile when boundaries matter: global_public, global_private, project_public, project_private, agent_session, and emergency_reconnect_archive. Discard is a disposition sink, not a durable scope.
  • MUST preserve source, status, scope, sensitivity, owner, evidence, and review metadata before retrieval treats a record as usable.
  • SHOULD add deterministic schema, redaction, dedupe, conflict, and source-quality checks before model-assisted curation.
  • SHOULD label user_identity, organization, and cross_organization memory only as namespaced optional extensions with stricter privacy rules.
  • MAY use centralized, federated, or hybrid topology when the deployment owner can justify privacy, availability, latency, and conflict tradeoffs.

Implementation Levels

Level Name Required artifacts Claim allowed
MATM-H1 Handbook-ready Roles, boundaries, taxonomy, scopes, source policy, and no-runtime boundary. The team has designed governed MATM.
MATM-H2 Proposal-and-curation-ready Memory Event schema, quarantine, curation reports, audit trail, direct-write prohibition. Workers can emit proposal-only Memory Events.
MATM-H3 Retrieval-ready Reviewed records indexed with scope, status, evidence, provenance, citations, and abstention rules. Consumers can retrieve reviewed scoped memory.
MATM-H4 Evaluated deployment Metrics for routing, leakage, evidence, redundancy, conflicts, utility, latency, cost, privacy, review burden, and adversarial behavior. The deployment has measured field behavior under its own assumptions.

Artifact Map

Artifact Level Route or bundle path
MATM_POLICY.md H1 llm-wiki/matm/MATM_POLICY.md
Memory Event schema H2 llm-wiki/matm/schemas/memory-event.schema.json
Curation report schema H2 llm-wiki/matm/schemas/curation-report.schema.json
Transactive directory schema H3 llm-wiki/matm/schemas/transactive-directory.schema.json
Retrieval request/response schemas H3 llm-wiki/matm/schemas/retrieval-*.schema.json
Evaluation plan H4 llm-wiki/matm/EVALUATION_PLAN.md

Non-Claims

Adopting this profile does not prove production readiness, security approval, privacy compliance, UAIX conformance, benchmark superiority, certification, or that LLMWikis operates the runtime. The deployment owner must implement, test, and approve its own system.