Multi-Agent Transactive Memory (MATM) is population-level memory for agents: producer agents contribute reviewable artifacts such as action-observation trajectories, and consumer agents retrieve relevant artifacts later to improve task work. The LLMWikis governed MATM profile adds curation, scopes, source status, privacy, and review gates so shared agent experience can become useful memory without letting workers write durable truth directly.
Public authority boundary
This page is handbook guidance. LLMWikis.org publishes schemas, templates, examples, and setup packets. It is not a live Memory Curator service, public MATM repository, write API, MCP server, A2A implementation, hosted vector database, event bus, agent runtime, automatic publication path, live evaluation service, certification program, or UAIX conformance authority.
Definition
MATM research core: Kim et al. describe producer agents that contribute agent-generated trajectories and consumer agents that retrieve relevant trajectory segments at inference time. The core empirical claim is about retrieval utility, not governance of durable organizational memory.
LLMWikis governed MATM profile: LLMWikis combines that trajectory-sharing idea with a separate Memory Curator governance design. Workers emit Memory Events into quarantine. A curator validates, redacts, classifies, deduplicates, checks conflicts, and chooses one of five durable disposition operations: append, update, deprecate, conflict, or discard.
| Term | Meaning in this profile |
|---|---|
| Producer agent | Agent that emits observable artifacts from a task, such as trajectory steps, tool results, tests, or concise decision rationales. |
| Consumer agent | Agent that retrieves reviewed records to help with a later task. |
| Agent-generated artifact | Reviewable output of an agent run: action-observation trajectory, plan, tool call, outcome summary, test result, or evidence reference. |
| Trajectory | Ordered action, observation, state, and result sequence from an agent task. It is evidence input, not automatically durable truth. |
| Transactive directory | Index that says which scoped records exist, who produced them, what they cover, status, provenance, trust signals, and retrieval rules. |
| Curator | Policy and review owner for durable memory admission. The curator can be deterministic, model-assisted, human-led, or hybrid. |
| Durable memory | Reviewed, scoped, retrievable memory record with source, status, owner, sensitivity, and audit metadata. |
| Working/session memory | Ephemeral task context that expires or is discarded unless explicitly promoted. |
| Retrieval planner | Component that chooses filters, scopes, ranking, abstention, and citation requirements before a consumer uses memory. |
| Provenance | Origin trail for event, artifact, source, agent identity, task, evidence references, and curation decision. |
| Trust signal | Review state, evidence quality, producer reliability, freshness, scope, conflict status, and sensitivity metadata used during retrieval. |
Reader-Action Map
| Reader goal | Start here | Next page |
|---|---|---|
| Understand the idea | This definition and source trace. | MATM reference profile |
| Decide whether a wiki should become MATM-style | Use ordinary LLM Wiki first, then add MATM only for shared reusable agent experience. | MATM Wikis guide |
| Add governed agent memory to an LLM Wiki | Use proposal-only Memory Events and curation reports. | Memory Events |
| Design an external MATM runtime | Separate worker execution, event queue, curator, durable store, and retrieval. | Curation |
| Migrate a direct-write system | Freeze direct writes, inventory stores, then run shadow curation. | Migration path |
| Evaluate a deployment | Separate research evidence, curator simulation, and deployment field evidence. | Evaluation |
| Review privacy or safety | Check scope leakage, memory poisoning, retention, and public-export review. | Security |
| Compare MATM with RAG or AI Memory | Use the comparison table below. | LLM Wiki vs RAG and AI Memory comparison |
Deployment Boundary Map
flowchart LR
handbook[LLMWikis handbook] -. guidance only .-> runtime[External MATM runtime]
worker[Worker / producer execution] --> event[Memory Event emission]
event --> queue[Quarantine or proposal queue]
queue --> validate[Deterministic validation]
validate --> redact[Safety and redaction]
redact --> classify[Curator classification]
classify --> review[Human or specialist review]
review --> durable[Durable scoped memory]
durable --> index[Index and transactive directory]
index --> retrieve[Retrieval planner and ranking]
retrieve --> consumer[Consumer execution]
durable --> audit[Audit and evaluation]
Text equivalent: The public handbook points to an external deployment. In that deployment, workers emit Memory Events; events wait in quarantine; deterministic checks and redaction run before curation; the curator classifies scope, kind, evidence, duplicates, and conflicts; human or specialist review gates durable admission; reviewed records are indexed; consumers retrieve with filters, citations, and abstention; audit and evaluation measure the loop. LLMWikis.org does not execute any of those runtime components.
Why MATM
- Agents repeatedly rediscover procedures, traps, state transitions, and environment-specific details that a later agent could reuse.
- Trajectories are often more useful than prose because they preserve what action was tried, what was observed, and what outcome followed.
- Population-level reuse separates producers from consumers, so one agent’s successful or failed work can help another without joint training.
- Uncontrolled writes create a governance problem: stale, private, poisoned, contradictory, or low-evidence memories can become highly retrievable.
MATM Versus Adjacent Systems
| System | Primary artifact | Write path | Retrieval behavior | Runtime ownership |
|---|---|---|---|---|
| Ordinary RAG | Chunks from documents or pages. | Indexing pipeline. | Similarity or hybrid retrieval over available corpus. | Retriever owner. |
| Single-agent memory | One agent’s preferences, notes, or task state. | Agent or app-specific memory update. | Personalized to that agent/session. | Agent app owner. |
| LLM Wiki durable knowledge | Reviewed human-readable pages with source status. | Source-governed page edits and review gates. | Readers route through canonical pages, metadata, and citations. | Wiki owner. |
| UAIX AI Memory or Project Handoff | Portable context packet for a bounded job or transfer. | Prepared local files and handoff rules. | Hot context for a receiver, linked back to durable sources where possible. | Project/handoff owner; UAIX is canonical for definitions. |
| Orchestration run state | Current task plan, tool state, retries, traces, and intermediate outputs. | Runtime orchestrator. | Used inside the current run. | Orchestrator owner. |
| MATM | Agent-generated trajectories and related artifacts. | Producer contribution and repository indexing. | Consumers retrieve relevant artifacts to improve task execution. | External MATM deployment owner. |
| Governed MATM | Reviewed Memory Records derived from Memory Events and evidence. | Proposal-only events, curator decisions, review, and audit. | Scope/status/evidence-aware retrieval with citations and abstention. | External deployment owner plus human stewards. |
Architecture Patterns
| Pattern | Consistency | Privacy boundary | Failure mode | Good fit |
|---|---|---|---|---|
| Centralized | One curator and directory decide durable memory. | Strong central policy, but broad trust in one store. | Single bottleneck or outage affects all consumers. | Small or medium team with common governance. |
| Federated/distributed | Each project or agent group curates local memory and shares selected records. | Better local privacy, harder cross-scope consistency. | Directory divergence, stale sync, or uneven policy. | Multiple tenants, domains, or regulatory boundaries. |
| Hybrid | Local curation plus shared directory for reviewed cross-scope records. | Local isolation with explicit promotion to shared memory. | Promotion latency or conflict between local and shared views. | Practical default for teams that need both isolation and reuse. |
Hybrid is practical guidance, not a universal proof. A deployment should choose curator placement, synchronization, conflict policy, and retrieval latency based on its own privacy, availability, and review constraints.
Roles
| Role | Owns | Forbidden action |
|---|---|---|
| Producer or worker | Observable task artifacts and Memory Event emission. | Writing durable memory directly. |
| Consumer | Using retrieved reviewed records in a later task. | Treating hypothesis or deprecated records as fact. |
| Memory Curator | Admission, scope, kind, evidence, dedupe, conflict, and disposition decisions. | Approving private or unsupported records without required review. |
| Policy/filter tier | Deterministic schema, redaction, allowlist, and scope checks. | Replacing human review for high-impact sensitive decisions. |
| Orchestrator | Run state, task routing, tools, retries, and trace capture. | Promoting run traces as durable truth automatically. |
| Specialist reviewer | Domain review for safety, privacy, legal, production, or evidence-sensitive records. | Expanding scope beyond the approved authority boundary. |
| Human steward | Governance, retention, escalation, and final publication decisions. | Claiming certification from handbook adoption. |
| Storage/indexing layer | Durable records, directory, search, retrieval cache, and citations. | Indexing discarded, private, or unreviewed events as current memory. |
| Audit/evaluation layer | Metrics, curation reports, leakage checks, latency, cost, and review burden. | Combining research, simulation, and field evidence into one unsupported claim. |
Lifecycle
Workers emit Memory Events. Events enter quarantine, pass schema and policy checks, are redacted where needed, receive scope and kind classification, attach evidence, pass duplicate and conflict analysis, receive a disposition, and require review before durable records become retrievable. Retrieval must preserve scope, status, source, citations, conflict state, and abstention rules.
Migration Path From Direct-Write Memory
- Freeze or gate direct durable writes.
- Inventory stores and writers.
- Classify sensitive, private, stale, and unsupported content.
- Separate session memory from durable memory.
- Backfill provenance where possible.
- Add Memory Event emission.
- Add proposal/quarantine storage.
- Run curator decisions in shadow mode.
- Measure routing, leakage, evidence, conflicts, latency, and downstream utility.
- Enable approved durable writes by curator only.
- Re-index only reviewed records.
- Add revalidation, deprecation, retention, audit, and incident cadence.
Implementation Checklist
- Run the MATM Wiki fit decision. Keep an ordinary LLM Wiki unless shared agent experience needs curated, scoped, retrievable, auditable memory.
- Pick an implementation level and name runtime owner.
- Adopt Memory Event schemas with four core scopes and nine core kinds.
- Install proposal-only curation before any durable writes.
- Define retrieval filtering, ranking, citation, and abstention rules.
- Review security, privacy, poisoning, and cross-scope leakage controls.
- Evaluate with separated evidence classes.
- Generate a MATM planning packet and stage it for review.
- Download the starter bundle and use the optional
llm-wiki/matm/add-on.
Source Trace
Research terms and benchmark context come from the MATM paper and repository. Governance scopes, curation, and write controls come from the separate Memory Curator design adapted as an LLMWikis profile. The two are intentionally separated.
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.