A MATM Wiki is an LLM Wiki that deliberately adds governed Multi-Agent Transactive Memory surfaces: Memory Events, curator decisions, scoped Memory Records, a transactive directory, retrieval policy, provenance receipts, adoption preview, rollback evidence, and evaluation. It is optional. Most teams should start with a regular source-governed LLM Wiki and add MATM only when agent experience reuse needs its own governance.
Fit rule
Do not use MATM because the name sounds advanced. Use MATM when multiple agents generate reusable operational experience and later agents need to retrieve that experience with source status, scope, permissions, citations, abstention, review gates, and rollback. If the work is mostly human documentation, static sources, or one-agent preference memory, a regular LLM Wiki is simpler and safer.
Ordinary LLM Wiki Or MATM Wiki?
| Question | Ordinary LLM Wiki | MATM Wiki |
|---|---|---|
| What is the main reusable thing? | Reviewed pages, source summaries, runbooks, policies, decisions, and evidence records. | Reviewed agent-generated trajectories, task outcomes, procedures, risks, conflicts, and reusable operational experience. |
| Who writes candidate material? | Humans and AI assistants draft pages or source summaries for review. | Producer agents emit proposal-only Memory Events; curators decide durable admission. |
| What is the risky failure? | Flat hierarchy, stale pages, weak source status, poor navigation, or unclear authority. | Cross-scope leakage, memory poisoning, blind imports, direct durable worker writes, and validation mistaken for trust. |
| What extra public surface is needed? | Source policy, metadata, index, log, sitemap, llms.txt, and agent rules. | All ordinary surfaces plus MATM profile, Memory Event schema, curation/retrieval policy, status labels, rollback/adoption rules, and non-claims. |
| When should it wait? | When sources are not inventoried or reviewed. | When curator ownership, privacy review, scope policy, or retrieval abstention rules are missing. |
Dogfooded Ecosystem Lanes
The current ecosystem split is the teaching pattern: NeuroWikis is human-facing education, NeuralWikis is the agent-facing exchange and MATM-style packet/profile layer, and LLMWikis is builder guidance for practical LLM Wiki and MATM Wiki implementation. That split prevents one page from trying to teach humans, serve agents, host schemas, and sell runtime claims at the same time.
| Lane | What belongs there | What does not belong there |
|---|---|---|
| Human education lane | Concepts, glossary, examples, diagrams, and route instructions for people. | Dense machine manifests, raw packet imports, private traces, or every source-memory route in global chrome. |
| Agent exchange lane | Public profiles, schema routes, llms.txt, routers, compatibility checks, status labels, proposal queues, and safety gates. | Implicit trust, credentials, direct durable writes, production authorization, or hidden private context. |
| Builder guidance lane | Decision guides, starter files, policy templates, implementation levels, tests, anti-patterns, and support boundaries. | Claims that the handbook operates the runtime, certifies systems, or owns UAIX conformance. |
This is also where the header/footer lesson lands: global chrome is a route chooser, not a memory inventory. Put complete source/site inventories behind directories, source maps, site-scoped memory guides, sitemap, and llms.txt.
Required MATM Wiki Artifacts
- Public route map:
llms.txt, sitemap, source map, and a small guide page that explains what agents should load first. - Machine-readable MATM profile with version, canonical routes, status labels, capabilities, source authority, support boundary, and explicit non-claims.
- Memory Event schema that treats worker output as proposal-only and excludes secrets, hidden reasoning, unsupported fields, and direct durable write claims.
- Curator runbook with validation, redaction, dedupe, scope routing, evidence checks, conflict preservation, disposition, audit, and human-review triggers.
- Retrieval policy that filters by scope, sensitivity, status, freshness, conflict state, authorization, citations, and abstention before ranking.
- Memory Firewall controls: no blind import, no private trace exposure, no cross-scope leakage, no validation-as-trust shortcut, and no retrieval-as-authorization shortcut.
- Adoption preview and rollback rule: packets or records must be sandboxed, source-traced, reversible, and reviewed before they affect durable state.
- Evaluation plan that separates MATM research evidence, curator simulation evidence, and deployment-specific field evidence.
Route Taxonomy From NeuralWikis Dogfood
A MATM Wiki needs a route taxonomy before it needs a larger navigation bar. The NeuralWikis implementation pattern keeps public discovery, non-mutating previews, protected write paths, reviewer decisions, operator evidence, and storage truth labels separate so agents do not confuse visibility with permission.
| Route class | Typical artifacts | Required boundary |
|---|---|---|
| Public GET discovery | llms.txt, router/profile JSON, capability registry, schema catalog, public-safe graph or wiki projection. |
Discovery is not credentials, trust, write permission, private workspace access, or production readiness. |
| Public POST preview | Schema validation, curation preview, retrieval preview, packet compatibility, adoption readiness. | Preview is non-mutating and advisory. Validation is not trust, and compatibility is not approval. |
| Protected submitter routes | Agent registration, Memory Event submit, trajectory submit, source ingest, wiki revision compile. | Require authorization, workspace scope, idempotency keys, provenance, redaction, and review-pending status. |
| Protected reviewer decisions | Approve, request changes, reject, quarantine, revoke, record feedback, review marketplace or packet listings. | Reviewer notes and private payloads stay redacted; decisions write audit evidence and never publish by existence alone. |
| Protected reader routes | Authorized retrieval query, private search, private ask, redacted graph index, workspace export manifests. | Retrieval grants no mutation or purchase authority and must not echo private questions, raw sources, hidden reasoning, or secrets. |
| Operator evidence routes | Observability, backup manifest, restore dry-run, maintenance run, outbox status, firewall reports. | Expose redacted evidence and configuration booleans only. Do not leak local paths, environment names or values, object keys, metric labels, raw payloads, or secret material. |
| Storage truth labels | Mode, durableConfigured, table families, fallback counts, worker commands, provider configuration status. | Memory fallback, local test adapters, and configured durable storage must be labeled separately before any production claim. |
When adapting this pattern, publish the route classes in the MATM profile and starter policy. Keep public pages concise, put dense endpoint inventories behind machine-readable files, and test that protected routes cannot be discovered as public authority through header/footer chrome.
Best Practices
| Practice | Reason |
|---|---|
| Start with the ordinary LLM Wiki contract. | Raw sources, reviewed pages, metadata, index/log, trust labels, and source policy are still the base. |
| Make the MATM acronym local and disambiguated. | MATM has collision risk; public pages should define it as Multi-Agent Transactive Memory on first use and avoid flooding unrelated human-facing pages. |
| Keep producer agents proposal-only. | Useful trajectories still need source, scope, evidence, privacy, conflict, and review checks before they become retrievable durable memory. |
| Publish status labels beside machine routes. | Agents need to know whether a route is profile, draft schema, protected endpoint, preview, non-claim, or live owned service. |
| Require citations and abstention in retrieval. | A retrieved record should explain what it is, where it came from, why it is usable, and when the consumer must stop. |
| Preserve conflicts as records. | Semantic averaging can erase important disagreement; conflicts need owners, evidence, and review state. |
| Keep public and protected routes separate. | Public profile/schema routes are not credentials, write permission, adoption authority, or production access. |
| Record rollback evidence before adoption. | MATM-style systems can change future agent behavior; reversal and quarantine must exist before promotion. |
Anti-Patterns
- Adding MATM to a small wiki where ordinary reviewed pages and retrieval rules are enough.
- Letting worker agents write durable Memory Records directly.
- Treating schema validation, compatibility preview, or HMAC attestation as trust by itself.
- Treating retrieval visibility as authorization to act.
- Flattening every source, site, schema, guide, and report into global header/footer navigation.
- Importing cognitive packets, trajectories, or external memory without sandbox, provenance, review, and rollback.
- Publishing private traces, hidden reasoning, credentials, regulated data, or raw user identity memory as durable public examples.
- Claiming live runtime, Memory Curator service, public write API, MCP/A2A support, vector database, certification, conformance, or production readiness from handbook copy.
Build Checklist
- Confirm the fit: shared reusable agent experience exists and ordinary LLM Wiki retrieval is insufficient.
- Name the lane split: human education, agent exchange, builder guidance, and canonical standards authority.
- Create or review
llm-wiki/matm/MATM_WIKI_DECISION_GUIDE.mdbefore enabling MATM mode. - Publish the route map, MATM profile, schemas, policy files, support boundaries, and discovery entries.
- Classify every endpoint or file as public discovery, public preview, protected submit, protected review, protected read, operator evidence, or storage-status evidence.
- Run curation in shadow mode before durable admission.
- Test invalid fixtures for direct worker writes, unsupported scopes, missing required fields, non-UTC timestamps, and secret-like fields.
- Verify public copy says validation is not trust, retrieval is not authorization, and profile visibility is not production access.
- Keep complete source inventories behind directories/source maps/discovery files instead of global chrome.
Source Trace
This guide is a builder-facing synthesis of the MATM research line, the separate Memory Curator governance pattern, and the current public NeuroWikis/NeuralWikis/LLMWikis lane split. It uses those sources as implementation guidance, not as a claim that LLMWikis operates a live MATM runtime.
MATM paper
Primary research source for producer and consumer agents sharing agent-generated trajectories through retrieval.
MATM repository
Reference implementation context and experiment README. Use the repository for direct checks of runtime details.
Memory Curator design
Separate governance design used as a practical control model for durable writes.
NeuralWikis MATM surface
Ecosystem dogfood for agent-facing MATM route boundaries, schemas, public profiles, review gates, and explicit non-claims.
NeuralWikis profile
Machine-readable profile pattern for capability status, truth labels, public/private boundaries, and no-runtime claims.
NeuroWikis human lane
Human-facing companion surface that keeps education separate from machine exchange and routes builders back to LLMWikis.
Transactive memory lineage
Original group-memory theory context for transactive memory terminology.