@mnemo-ai/mcp

AI pointer: Standalone MCP stdio server. Proxies tool calls to the HTTP API via @mnemo-ai/client-ts. 34 tools: all 34 live (all are registered and functional in the server; see table below). Tool count is enforced by registerAllTools/MCP_TOOL_COUNT in src/registry.ts — treat that file as the source of truth, not this README, if they ever disagree.

Depends on: @mnemo-ai/client-ts, @modelcontextprotocol/sdk

Live tools (34 registered): memory_recall, memory_remember, memory_pin, memory_forget, memory_timeline, memory_search, memory_relate, memory_decide, memory_entity, memory_status, memory_remind_when, memory_consolidate, memory_session, memory_episode_start, memory_episode_end, memory_task, memory_event, memory_reasoning, memory_strategy, memory_self_edit, memory_hyperedge, memory_introspect, memory_anticipate, memory_compose, memory_workflow, memory_skill, memory_doctor, memory_update, memory_judge, memory_stats, memory_compare, memory_purge, memory_context, memory_workspace

All 34 tools are live. There are no stub tools.

Status: BETA (all 34 live)

See also: CLAUDE.md · docs/INDEX.md · docs/STATUS.md

Standalone Model Context Protocol server for Mnemosyne — the multi-tenant cognitive memory system.

Plug it into any MCP client (Claude Desktop, Cursor, Continue, …) and your assistant gets durable, workspace-scoped memory: recall what users prefer, remember decisions across sessions, pin critical knowledge, and audit what was learned over time.

This package is a thin adapter. It speaks MCP over stdio and forwards every call to @mnemo-ai/client-ts, which in turn talks to a Mnemosyne HTTP server you control.

Install

You don't need to install anything globally. Just point your MCP client at it via npx.

Configure

Claude Desktop

Add this to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "mnemosyne": {
      "command": "npx",
      "args": ["-y", "@mnemo-ai/mcp"],
      "env": {
        "MNEMO_URL": "https://memory.acme.com",
        "MNEMO_KEY": "msk_..."
      }
    }
  }
}

Restart Claude Desktop. You should see "mnemosyne" appear in the MCP tools menu.

Cursor

In ~/.cursor/mcp.json (or the per-project equivalent):

{
  "mcpServers": {
    "mnemosyne": {
      "command": "npx",
      "args": ["-y", "@mnemo-ai/mcp"],
      "env": {
        "MNEMO_URL": "https://memory.acme.com",
        "MNEMO_KEY": "msk_..."
      }
    }
  }
}

Continue / any other MCP client

Same pattern: command: npx, args: ["-y", "@mnemo-ai/mcp"], set MNEMO_URL and MNEMO_KEY in env.

Environment variables

| Var | Required | Description | | --- | --- | --- | | MNEMO_URL | yes | Base URL of your Mnemosyne server, e.g. https://memory.acme.com. | | MNEMO_KEY | yes | Workspace API key (msk_... or mk_live_...). Scope determines which tools are available — read-only keys cannot call remember, pin, or forget. |

If either is missing, the server exits with code 1 and prints a help message to stderr (which most MCP clients surface in their UI).

Tools exposed

Status legend: a Live tool is registered and functional. All 34 tools are live.

| Tool | Status | Scope | Purpose | | --- | --- | --- | --- | | memory_recall | ✅ Live | read | Top-K relevant facts for a query (semantic + worth-weighted). | | memory_remember | ✅ Live | write | Persist a new durable fact (idempotent via content hash). | | memory_pin | ✅ Live | write | Protect a fact from auto-prune and boost it in recall. | | memory_forget | ✅ Live | write | Soft-archive a fact (closes validTo, auditable, restorable). | | memory_timeline | ✅ Live | read | Fact changes over a time window ("what did my AI learn this week?"). | | memory_search | ✅ Live | read | Full-text + semantic search with filters (kind, sensitivity, date). | | memory_relate | ✅ Live | write | Create a relation between two primitives (9 locked verbs). | | memory_decide | ✅ Live | write | Log an architectural or policy decision with full rationale. | | memory_entity | ✅ Live | write | Upsert a named entity (person, org, project, concept). | | memory_status | ✅ Live | read | Health check: server version, adapter states. | | memory_remind_when | ✅ Live | write | Schedule a prospective memory trigger on a condition or time. | | memory_consolidate | ✅ Live | write | Trigger a consolidation pass (merge/retire duplicate facts). | | memory_session | ✅ Live | write | Unified session control (§24): pin a fact, append a scratch turn, collapse, or discard. All 4 actions fully wired. | | memory_episode_start | ✅ Live | write | Open a named episode context (session). | | memory_episode_end | ✅ Live | write | Close and collapse an episode, extracting facts from transcript. | | memory_task | ✅ Live | write | Gestiona tareas accionables: crear, leer, listar, actualizar o archivar con dueño, estado y dependencias. | | memory_event | ✅ Live | write | Registra eventos inmutables en el log del workspace: create, get, list, archive. | | memory_reasoning | ✅ Live | write | Persiste chains-of-thought estructurados (pregunta → pasos → conclusión) reutilizables. | | memory_strategy | ✅ Live | write | Gestiona estrategias reusables trigger→action. Usa op como discriminador (no action). | | memory_self_edit | ✅ Live | write | Corrige el statement de un fact existente con trazabilidad: guarda la versión previa y el motivo en el audit log. | | memory_hyperedge | ✅ Live | write | Create/relate an N-ary hyperedge across ≥3 facts (hypergraph associations beyond binary relations). | | memory_introspect | ✅ Live | read | Self-model / meta-cognition: knowledge coverage, gaps, and confidence trends per entity. | | memory_anticipate | ✅ Live | read | Predictive next-recall suggestions from temporal + co-occurrence patterns (anticipation engine). | | memory_compose | ✅ Live | read | Compose new inferences over recalled facts via the declarative inference engine. | | memory_workflow | ✅ Live | write | Manage reusable workflow definitions: create, get, list, patch, archive. | | memory_skill | ✅ Live | write | Manage reusable skill definitions: create, get, list, patch, archive. | | memory_doctor | ✅ Live | read | Runs 8 operational diagnostics (Postgres connectivity, app_user grants, schema shadows, migration version, embedding backlog, workspace_id validity, MCP env config, tool registration count). No input. Use before filing a bug report. | | memory_update | ✅ Live | write | Corrige el statement de un fact existente vía reemplazo bitemporal: cierra el fact original y crea uno nuevo con el mismo linaje. El id del fact cambia. | | memory_judge | ✅ Live | write | Resuelve conflictos entre facts contradictorios. action='list' muestra pares en conflicto; action='resolve' aplica supersede_a | supersede_b | keep_both | merge. | | memory_stats | ✅ Live | read | Estadísticas agregadas del workspace: facts live/closed/pinned/archived, decisiones activas, conflictos, cobertura de embeddings, hit-rate de recall a 30 días. Sin input. | | memory_compare | ✅ Live | read | Compara dos facts por ID: contenido completo, diffs de campos, verbo de relación sugerido (supersedes | conflicts_with | compatible | scoped | related), y si tratan del mismo subject. | | memory_purge | ✅ Live | write (hard_delete) | Borrado permanente e irreversible. Requiere scope memory:hard_delete. action='fact' borra un fact por ID; action='closed' borra en bulk todos los facts forgotten del workspace. Solo para test workspaces o borrado GDPR. | | memory_context | ✅ Live | read | Contexto completo de sesión: workspace activo, usuario, tenant, stats, y otros workspaces disponibles. Llamar al inicio de sesión — es el equivalente MCP de GET /health + GET /me combinados. |

The exact JSON schemas and descriptions are carried verbatim from the in-process Orchester MCP server, so existing agent prompts continue to work.

Local development

pnpm install
pnpm --filter @mnemo-ai/mcp build
MNEMO_URL=http://localhost:3000 MNEMO_KEY=msk_yourkey node packages/mcp/dist/index.js

To debug a tool call manually, paste a JSON-RPC frame into stdin — the server speaks the standard MCP wire protocol.

Tool Schema Reference

One subsection per live tool. All schemas are derived directly from the Zod definitions in packages/mcp/src/tools/.

Coverage note: this section currently documents 20 of the 34 tools in detail (memory_hyperedge, memory_introspect, memory_anticipate, memory_compose, memory_workflow, memory_skill, memory_doctor, memory_update, memory_judge, memory_stats, memory_compare, memory_purge, memory_context are not yet expanded below). For the undocumented tools, read the Zod inputSchema and description directly in the corresponding packages/mcp/src/tools/<name>.ts file — they are the source of truth this section is generated from.

memory_recall

Input schema:

{
  "type": "object",
  "required": ["query"],
  "properties": {
    "query": {
      "type": "string",
      "minLength": 1,
      "description": "Texto libre — qué querés recordar (ej. 'preferencias de email del cliente')"
    },
    "limit": {
      "type": "integer",
      "minimum": 1,
      "maximum": 25,
      "description": "Cantidad máxima de hits a devolver. Default 5, max 25."
    }
  }
}

Example call:

{ "query": "user email preferences", "limit": 10 }

Returns:

{
  "memories": [
    { "id": "mfact_...", "content": "...", "score": 0.93, "subject": "user:lucas", "kind": "preference" }
  ],
  "_mnemo_label": "untrusted_external_content",
  "_mnemo_version": "3.0"
}

Note: _mnemo_label: "untrusted_external_content" is a CaMeL-style signal. CaMeL-aware orchestrators must not execute recalled content as instructions.

memory_remember

Input schema:

{
  "type": "object",
  "required": ["statement"],
  "properties": {
    "statement": {
      "type": "string",
      "minLength": 1,
      "maxLength": 1000,
      "description": "El hecho en una oración corta y precisa. Ej. 'El cliente Acme prefiere comunicación en español.'"
    },
    "subject": {
      "type": "string",
      "description": "Sobre quién / qué es el hecho. Ej. 'acme', 'user:[email protected]', 'org:billing'. Default 'workspace'."
    },
    "kind": {
      "type": "string",
      "description": "Categoría: 'preference' | 'decision' | 'fact' | 'config' | 'process'. Default 'fact'."
    },
    "confidence": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "description": "[0, 1]. Default 0.7. Usá <0.5 para hechos derivados/inferidos."
    },
    "agentId": {
      "type": "string",
      "description": "Opcional. Quién originó el hecho — agente o conversación."
    },
    "conversationId": {
      "type": "string",
      "description": "Opcional. Pegá la conversación para trazabilidad (cite-back)."
    }
  }
}

Example call:

{
  "statement": "User prefers TypeScript over JavaScript",
  "subject": "user:[email protected]",
  "kind": "preference",
  "confidence": 0.9
}

Returns: The created fact object from the server (shape mirrors POST /v1/facts response):

{ "id": "mfact_...", "content": "User prefers TypeScript over JavaScript", "createdAt": "2026-06-13T10:00:00Z" }

memory_pin

Input schema:

{
  "type": "object",
  "required": ["factId"],
  "properties": {
    "factId": {
      "type": "string",
      "minLength": 1,
      "description": "ID del fact a pinear."
    },
    "unpin": {
      "type": "boolean",
      "description": "Pass true to UNPIN the fact instead of pinning it. Default false."
    }
  }
}

Example call:

{ "factId": "mfact_01j..." }

Example unpin call:

{ "factId": "mfact_01j...", "unpin": true }

Returns: The updated fact object from the server (shape mirrors POST /v1/facts/:id/pin response).

memory_forget

Input schema:

{
  "type": "object",
  "required": ["factId"],
  "properties": {
    "factId": {
      "type": "string",
      "minLength": 1,
      "description": "ID del fact a archivar."
    },
    "reason": {
      "type": "string",
      "description": "Motivo opcional (queda en audit log)."
    }
  }
}

Example call:

{ "factId": "mfact_01j...", "reason": "Information was incorrect" }

Returns:

{ "id": "mfact_01j...", "forgotten": true, "reason": "Information was incorrect" }

Note: This is a soft delete. The fact moves to mnemo_fact_archive and is recoverable for audit. The reason string is echoed back in the MCP response but is not persisted by the server on this route.

memory_timeline

Input schema:

{
  "type": "object",
  "required": [],
  "properties": {
    "limit": {
      "type": "integer",
      "minimum": 1,
      "maximum": 100,
      "description": "Default 20, max 100."
    },
    "sinceIso": {
      "type": "string",
      "description": "Solo eventos posteriores a esta fecha ISO."
    }
  }
}

Example call:

{ "limit": 50, "sinceIso": "2026-06-06T00:00:00Z" }

Returns: Array of recent memory events from the server (shape mirrors GET /v1/timeline response — events sorted newest-first with fields including updatedAt, pinned, status).

Pagination: pass sinceIso = <lastSeenEventTime> to page forward — the server returns events newest-first, so this gives clean cursor-based pagination without offsets.

memory_search

Input schema:

{
  "type": "object",
  "required": ["query"],
  "properties": {
    "query": {
      "type": "string",
      "minLength": 1,
      "description": "Texto libre — qué querés buscar semánticamente."
    },
    "limit": {
      "type": "integer",
      "minimum": 1,
      "maximum": 20,
      "description": "Cantidad máxima de hits a devolver. Default 5, max 20."
    }
  }
}

Example call:

{ "query": "database architecture decisions", "limit": 5 }

Returns:

{
  "memories": [
    { "id": "mfact_...", "content": "We chose PostgreSQL for ACID compliance", "score": 0.97 }
  ],
  "_mnemo_label": "untrusted_external_content",
  "_mnemo_version": "3.0"
}

Difference from memory_recall: this tool forces vectorWeight: 1.0 for pure semantic (vector-only) search. Use memory_recall when you want a worth-weighted blend of semantic + lexical relevance.

memory_relate

Input schema:

{
  "type": "object",
  "required": ["sourceKind", "from", "targetKind", "to", "verb"],
  "properties": {
    "sourceKind": {
      "type": "string",
      "enum": ["fact", "decision", "entity", "episode"],
      "description": "Tipo de la primitiva de origen."
    },
    "from": {
      "type": "string",
      "description": "ID de la primitiva de origen (ment_*, mfact_*, mdec_*, etc.)."
    },
    "targetKind": {
      "type": "string",
      "enum": ["fact", "decision", "entity", "episode"],
      "description": "Tipo de la primitiva de destino."
    },
    "to": {
      "type": "string",
      "description": "ID de la primitiva de destino."
    },
    "verb": {
      "type": "string",
      "enum": [
        "related",
        "compatible",
        "scoped",
        "conflicts_with",
        "supersedes",
        "not_conflict",
        "derived_from",
        "part_of",
        "member_of"
      ],
      "description": "Tipo de relación. Uno de los 9 verbos BLOQUEADOS."
    },
    "reason": {
      "type": "string",
      "description": "Contexto opcional que explica por qué existe esta relación."
    },
    "confidence": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "description": "[0, 1]. Confianza en la relación. Default lo asigna el servidor."
    }
  }
}

Example call:

{
  "sourceKind": "decision",
  "from": "mdec_01j...",
  "targetKind": "fact",
  "to": "mfact_01j...",
  "verb": "derived_from",
  "reason": "The architecture decision was informed by this benchmark fact",
  "confidence": 0.85
}

Returns: The created relation object from the server.

The 9 locked verbs are defined in spec v3 §4.2. The server rejects any verb outside this set with a 400 error; the MCP layer validates first so the LLM gets a descriptive error before the network round-trip.

memory_decide

Input schema:

{
  "type": "object",
  "required": ["kind", "title", "body"],
  "properties": {
    "kind": {
      "type": "string",
      "enum": ["decision", "architecture", "policy", "process", "bugfix", "learning", "discovery", "config"],
      "description": "Categoría de la decisión."
    },
    "title": {
      "type": "string",
      "minLength": 1,
      "maxLength": 500,
      "description": "La decisión en una oración concisa. Ej. 'Usar PostgreSQL como base de datos principal.'"
    },
    "body": {
      "type": "string",
      "minLength": 1,
      "description": "Por qué se tomó esta decisión. Incluí el razonamiento, las alternativas consideradas y los factores determinantes."
    },
    "topicKey": {
      "type": "string",
      "description": "Clave normalizada para deduplicación (upsert). Omitir deja al servidor calcularla a partir de title."
    },
    "conversationId": {
      "type": "string",
      "description": "ID de la conversación en que se tomó la decisión."
    }
  }
}

Example call:

{
  "kind": "architecture",
  "title": "Use PostgreSQL as the primary database",
  "body": "We chose PostgreSQL over MySQL for its JSONB support, full ACID compliance, and better support for complex queries. MySQL was considered but ruled out due to weaker partial-index support.",
  "conversationId": "conv_01j..."
}

Returns: The created or upserted decision object from the server (stored in mnemo_decision, not mnemo_fact).

Idempotent by normalized topic_key. Re-submitting the same title (normalized) performs an upsert and returns the existing row. Pass an explicit topicKey to control the dedup key.

memory_entity

Input schema:

{
  "type": "object",
  "required": ["name", "kind"],
  "properties": {
    "name": {
      "type": "string",
      "minLength": 1,
      "maxLength": 200,
      "description": "Nombre canónico de la entidad. Ej. 'Acme Corp', 'Lucas Mailland', 'sistema-auth'."
    },
    "kind": {
      "type": "string",
      "enum": ["person", "organization", "project", "concept", "place", "other"],
      "description": "Categoría de la entidad."
    },
    "description": {
      "type": "string",
      "description": "Descripción breve: rol, contexto o detalles relevantes."
    },
    "aliases": {
      "type": "array",
      "items": { "type": "string" },
      "description": "Nombres alternativos o apodos."
    },
    "metadata": {
      "type": "object",
      "additionalProperties": true,
      "description": "Metadatos arbitrarios."
    }
  }
}

Example call:

{
  "name": "Acme Corp",
  "kind": "organization",
  "description": "Enterprise customer using the ATS product",
  "aliases": ["Acme", "acme-corp"],
  "metadata": { "plan": "enterprise", "region": "us-east" }
}

Returns: The created or retrieved entity object from the server (upsert by normalized name within the workspace).

memory_status

Input schema:

{
  "type": "object",
  "required": [],
  "properties": {}
}

Example call:

{}

Returns: Health snapshot from GET /v1/health, including sub-component status for graceful degradation:

{
  "status": "ok",
  "db": "ok",
  "embedder": "ok",
  "queue": "ok",
  "version": "3.0.0",
  "recallLatencyP50Ms": 42
}

Sub-component statuses allow orchestrators to degrade gracefully — e.g. if embedder is down but db is "ok", fact retrieval by ID may still work while semantic recall is unavailable.

memory_remind_when

Description: Schedule a prospective memory trigger. The server evaluates the condition against new incoming facts and fires the reminder when matched.

Input schema:

{
  "condition": "string — natural-language condition to watch for (required)",
  "label": "string — human-readable label for this reminder (required)",
  "triggerAt": "string — ISO-8601 datetime for time-based trigger (optional)",
  "metadata": "object — arbitrary caller metadata (optional)"
}

Output schema:

{
  "content": [{ "type": "text", "text": "Reminder scheduled: {label} (id: {id})" }]
}

Example call:

{
  "name": "memory_remind_when",
  "arguments": {
    "condition": "user mentions they want to switch to TypeScript",
    "label": "TS migration intent",
    "triggerAt": "2026-07-01T09:00:00Z"
  }
}

memory_consolidate

Description: Trigger a fast consolidation pass for the active workspace. The engine selects bounded low-worth candidate facts, merges eligible near-duplicate clusters, and reports candidate, merge, and skip counts. Only mode: "fast" with dryRun: false is currently supported. Manual deep mode returns HTTP 422 deep_mode_not_supported, and dryRun: true returns HTTP 422 dry_run_not_supported without running consolidation.

Input schema:

{
  "mode": "\"fast\" | \"deep\" - only fast is currently supported; deep returns HTTP 422 (optional, default: fast)",
  "dryRun": "boolean - must currently be false; true returns HTTP 422 (optional, default: false)"
}

Output schema:

{
  "content": [{
    "type": "text",
    "text": "Consolidation complete. candidates=N merged=M skipped=K"
  }]
}

Example call:

{
  "name": "memory_consolidate",
  "arguments": { "mode": "fast", "dryRun": false }
}

memory_session

Description: Unified session-control tool (spec §24). Dispatches on action: pin/unpin a fact, append a scratch turn to a session, collapse a session (triggering the extraction pipeline), or discard a session. The MCP layer is a thin adapter — the server owns the logic. Action → transport mapping:

• pin → client.pinFact / client.unpinFact (POST /v1/facts/:id/pin) — live
• collapse → client.collapseSession (POST /v1/session/:id/collapse) — live
• scratch → client.updateSessionScratch (POST /v1/session/:id/scratch) — live
• discard → client.discardSession (POST /v1/session/:id/discard) — live

Input schema:

{
  "action": "\"pin\" | \"scratch\" | \"collapse\" | \"discard\" — required",
  "sessionId": "string — required for collapse/scratch/discard",
  "factId": "string — required for pin",
  "unpin": "boolean — for pin (true = unpin; default false)",
  "content": "string — required for scratch (the turn text)",
  "role": "\"user\" | \"assistant\" | \"system\" — for scratch (turn author)",
  "collapseHint": "string — optional natural-language hint for collapse",
  "metadata": "object — optional",
  "idempotencyKey": "string — optional (forwarded as the Idempotency-Key header)"
}

Output schema:

{
  "content": [{ "type": "text", "text": "<result summary>" }],
  "structuredContent": { "<action-specific result>" }
}

All actions return a successful response with action-specific structured content (e.g., { "scratched": true, "sessionId": "..." } for scratch, { "discarded": true, "sessionId": "..." } for discard).

Example call:

{
  "name": "memory_session",
  "arguments": { "action": "pin", "factId": "fct_01hwz..." }
}

memory_episode_start

Note (spec §24 MCP Tools): The unified memory_session tool (above) now exposes the spec's action: pin | scratch | collapse | discard contract. pin and collapse are fully wired; scratch and discard return an actionable error pending server/client wiring (POST /v1/session/:id/turn and a discard endpoint). memory_pin, memory_episode_start, and memory_episode_end remain as focused aliases for the pin / open-session / collapse paths and continue to work.

Description: Open a named episode (session). Returns a sessionId that should be passed to subsequent memory_remember calls to group facts under the episode context. Close with memory_episode_end to trigger transcript extraction.

Input schema:

{
  "agentId": "string — identifier for the agent opening the episode (optional)",
  "actorId": "string — identifier for the human actor (optional)",
  "metadata": "object — arbitrary session metadata (optional)"
}

Output schema:

{
  "content": [{ "type": "text", "text": "Episode started. sessionId: {id}" }]
}

Example call:

{
  "name": "memory_episode_start",
  "arguments": { "agentId": "claude-sonnet", "actorId": "user-42" }
}

memory_episode_end

Description: Close an episode and trigger the 8-step collapse pipeline: the session transcript is sent to the LLM for fact extraction, which creates new facts tagged with the episode context. The session status changes to collapsed.

Input schema:

{
  "sessionId": "string — ID returned by memory_episode_start (required)",
  "collapseHint": "string — optional natural-language hint about what was important in this episode (optional)"
}

Output schema:

{
  "content": [{
    "type": "text",
    "text": "Episode collapsed. sessionId: {id} status: collapsed"
  }]
}

Example call:

{
  "name": "memory_episode_end",
  "arguments": {
    "sessionId": "ses_01hwz...",
    "collapseHint": "Focus on architecture decisions"
  }
}

memory_task

Description: Gestiona tareas accionables del workspace. Las tareas son trabajos puntuales con dueño, estado y dependencias. Verbos de dominio §24: delegate reasigna la tarea a otro agente (requiere assignedTo; aplica status → delegated); complete la marca como terminada (acepta outcome y lessonsLearned).

Input schema:

{
  "action": "\"create\" | \"get\" | \"list\" | \"patch\" | \"archive\" | \"delegate\" | \"complete\" — required",
  "id": "string — required for get/patch/archive/delegate/complete",
  "title": "string (max 500) — required for create",
  "description": "string (max 20000) — optional",
  "status": "\"pending\" | \"in_progress\" | \"blocked\" | \"delegated\" | \"completed\" | \"cancelled\" — optional",
  "priority": "\"critical\" | \"high\" | \"medium\" | \"low\" — optional",
  "assignedTo": "string — required for create",
  "sessionId": "string — optional",
  "parentTaskId": "string — optional",
  "dueAt": "ISO datetime — optional",
  "outcome": "string (max 20000) — optional",
  "lessonsLearned": "string (max 20000) — optional",
  "metadata": "object — optional",
  "parentTaskIdFilter": "string — list filter for parentTaskId",
  "limit": "integer (1–500) — optional"
}

Example call:

{
  "name": "memory_task",
  "arguments": {
    "action": "create",
    "title": "Migrate auth to Better Auth",
    "assignedTo": "agent-claude",
    "priority": "high"
  }
}

memory_event

Description: Registra eventos inmutables en el log del workspace. Los eventos son append-only; usá memory_fact para estado mutable.

Input schema:

{
  "action": "\"create\" | \"get\" | \"list\" | \"archive\" — required",
  "id": "string — required for get/archive",
  "type": "\"system\" | \"external\" | \"user\" | \"agent\" | \"scheduled\" — required for create",
  "name": "string (max 500) — required for create",
  "description": "string (max 20000) — required for create",
  "occurredAt": "ISO datetime — required for create",
  "source": "string (max 200) — required for create",
  "durationS": "integer ≥ 0 — optional",
  "relatedEntities": "string[] — optional",
  "metadata": "object — optional",
  "sourceFilter": "string — list filter for source",
  "after": "ISO datetime — list filter",
  "before": "ISO datetime — list filter",
  "limit": "integer (1–500) — optional"
}

Example call:

{
  "name": "memory_event",
  "arguments": {
    "action": "create",
    "type": "agent",
    "name": "wave-b-deployed",
    "description": "Wave B MCP tools deployed successfully",
    "occurredAt": "2026-06-15T01:00:00.000Z",
    "source": "mcp-agent"
  }
}

memory_reasoning

Description: Almacena chains-of-thought estructurados (pregunta → pasos → conclusión) para reutilizarlos en situaciones similares.

Input schema:

{
  "action": "\"create\" | \"get\" | \"list\" | \"patch\" | \"archive\" — required",
  "id": "string — required for get/patch/archive",
  "agentId": "string — optional",
  "question": "string (max 20000) — required for create",
  "chain": "array of { step: int, reasoning: string, conclusion?: string } (max 200 steps) — required for create",
  "conclusion": "string (max 20000) — required for create",
  "topicKey": "string (max 200) — optional",
  "status": "\"active\" | \"archived\" | \"pruned\" — optional",
  "metadata": "object — optional",
  "topicKeyFilter": "string — list filter for topicKey",
  "limit": "integer (1–500) — optional"
}

Example call:

{
  "name": "memory_reasoning",
  "arguments": {
    "action": "create",
    "question": "Should we use PostgreSQL or MongoDB for this use case?",
    "chain": [
      { "step": 1, "reasoning": "Need ACID transactions and relational joins" },
      { "step": 2, "reasoning": "Schema is stable and well-defined" }
    ],
    "conclusion": "PostgreSQL is the right choice"
  }
}

memory_strategy

Description: Gestiona estrategias reusables: heurísticas trigger→action que el agente aplica cuando ocurre una condición disparadora. Usa op como discriminador (no action, que es un campo de contenido). Verbos de dominio §24: store guarda una nueva estrategia (alias de create); recall recupera estrategias relevantes/activas (alias de list); rate registra el resultado (éxito/fallo) de una estrategia para ajustar su worth.

Input schema:

{
  "op": "\"create\" | \"get\" | \"list\" | \"patch\" | \"archive\" | \"store\" | \"recall\" | \"rate\" — required",
  "id": "string — required for get/patch/archive",
  "trigger": "string (max 20000) — required for create. La condición disparadora.",
  "action": "string (max 20000) — required for create. La heurística a aplicar.",
  "context": "string (max 20000) — optional",
  "agentId": "string — optional",
  "supersedesId": "string — optional",
  "status": "\"active\" | \"superseded\" | \"archived\" — optional",
  "metadata": "object — optional",
  "limit": "integer (1–500) — optional"
}

Example call:

{
  "name": "memory_strategy",
  "arguments": {
    "op": "create",
    "trigger": "API returns 5xx three times in a row",
    "action": "Switch to exponential backoff with jitter; alert after 5 minutes",
    "context": "execution"
  }
}

memory_self_edit

Description: Edita memoria propia con append, replace, strengthen, weaken, forget, reflect o link. Cuando replace apunta a un fact, el servidor cierra la versión histórica, crea un successor con ID nuevo y devuelve replacementId. El motivo es contexto obligatorio de política, pero todavía no se persiste como auditoría durable.

Input schema:

{
  "action": "append | replace | strengthen | weaken | forget | reflect | link — required",
  "target": "fact_id | profile | strategy:{id} — required",
  "content": "string (max 20000) — required for append/replace",
  "reason": "string (max 2000) — required policy context",
  "idempotencyKey": "string — optional"
}

Example call:

{
  "name": "memory_self_edit",
  "arguments": {
    "action": "replace",
    "target": "mfact_01j...",
    "content": "The user prefers dark mode in all applications",
    "reason": "Previous statement only mentioned VS Code; confirmed applies globally"
  }
}

Returns: A structured self-edit result. Applied fact replacements include targetKind: "fact" and the required new replacementId. Approval-required results do not include a replacement ID. Returns isError: true with a descriptive message on failure.

License

Apache-2.0 © Lucas Mailland