> Full SurrealDB documentation index: https://surrealdb.com/docs/llms.txt # MCP tools Tool payloads and ACL alignment. SurrealDB Agent Memory's MCP server exposes **seven** tools at `/mcp`: `remember`, `recall`, `context`, `reflect`, `forget`, `upload`, and `inspect`. Each maps to one or more REST endpoints and uses the same authentication and scope semantics. > [!NOTE] > Older docs and some third-party snippets used names like `memory_store` or `knowledge_search`. Those prefixes are gone - use the short names above. > [!NOTE] > REST responses use **camelCase** (`queryMs`, `traceId`, `trace.traceId`). **Scope selectors** on the wire use DNF (`ScopeSets`): writes take **`scope`** / **`scopes`**, reads take **`lens`** - an OR of conjunctive slash-path clauses. Register paths with `agent-memory scopes create` before use. See [Contexts and scope](/docs/agent-memory/mental-model/contexts-and-scope.md#wire-format-scopesets). ## Authentication All tools use **`Authorization: Bearer`** authentication (same as REST). Pass the token in the MCP configuration: ```json { "mcpServers": { "agent-memory": { "url": "https://spectron.example.com/mcp", "headers": { "Authorization": "Bearer sp-..." } } } } ``` Each API key is bound to exactly one Context, so **`context_id` is optional** on every tool: when omitted, the server resolves the Context from the bearer key. You may still pass `context_id` explicitly; a value that does not match the key's bound Context is rejected with **`401`** (masked, not `404`). ## Scope handling Scope is layered on each tool call: 1. **Grant region** (from the API key's principal) - enforced server-side; cannot be widened past what the key allows. 2. **Per-call `scope` / `lens` argument** - DNF selector narrowing the operation within the grant (`scope`/`scopes` on writes, `lens` on `recall` / `context`). Register paths with `agent-memory scopes create` before first use. --- ## `remember` Persist a new fact from text. Reconciles against existing memory (supersession, uncertainty). **Requires:** `memory:write` over the target scope. **Input:** ```jsonc { "text": "I just got promoted to CTO.", // required "session_id": "sess_abc123", // optional - creates a session if omitted "scope": [["org/acme/user/alice"]], // optional DNF write selector "labels": ["source=onboarding"], // optional "infer": "full", // full | triples | preview | none (default full) "observed_at": "2026-05-12T10:00:00Z" // optional, full and triples only } ``` `text` is required unless `infer` is `triples`. With `triples`, pass a **`triples`** array of structured entries instead, and no LLM runs. Each entry has an `entity` (`type` and `name`) and at most one of: `value` with a `key` (an attribute), `target` with a `key` (a relation), or `verb` (a dated action). An entry with none of these upserts the entity only. `preview` extracts but does not persist. **`observed_at`** (RFC 3339) sets the known time of the written facts. **Output:** `mode`, `session_id`, `turn_id`, `chunk_id` (for `none`), `extraction` (the structured diff of entities, attributes, relations, instructions, uncertainties and corrections), `preview` (for `preview`), and a stand-in `trace_id` (the facts path returns the turn id today). **Underlying REST endpoint:** `POST /api/v1/{context_id}/facts` --- ## `recall` Search the unified substrate (experiential facts + document passages) and return ranked hits. **Requires:** `memory:read`. **Input:** ```jsonc { "query": "What does Alice do at Acme?", // required "k": 10, // optional, default 10 (max 50) "mode": "hybrid", // vector | bm25 | graph | hybrid "lens": [["org/acme"]], // optional DNF read lens "labels": ["project=support"] // optional key=value filter } ``` **Output:** Ranked hits with scores and source kinds, plus `trace_id`. **Underlying REST endpoint:** `POST /api/v1/{context_id}/query` --- ## `context` Assemble a markdown context block (profile + preferences + relevant facts) ready for system-prompt injection. **Requires:** `memory:read`. **Input:** ```jsonc { "query": "What does Alice do at Acme?", "lens": [["org/acme"]], "labels": ["org=acme"] } ``` **Output:** Markdown context string plus `trace_id`. **Underlying REST endpoint:** `POST /api/v1/{context_id}/context` --- ## `reflect` Synthesise patterns across stored memory. With `persist: true`, writes the insight as new facts (`source.kind = "reflect"`). **Requires:** `memory:read`; `persist: true` also needs write access in the caller's region. **Input:** ```jsonc { "query": "What recurring complaints have customers raised this month?", "persist": true // optional, default false } ``` **Output:** Reflection text, evidence refs, optional `persisted_attributes`, and `trace_id`. **Underlying REST endpoint:** `POST /api/v1/{context_id}/reflect` --- ## `forget` Stop believing something. Soft-deletes the memories that match a natural-language query, whatever table they are in: retrieval stops serving them under current and temporal reads, but the records stay stored, and a system-time (`atInstant`) read anchored before the forget can still see them. `purge: true` is the irreversible right-to-be-forgotten erasure that also removes that stored history. **Requires:** `memory:forget` (not the same as `memory:write`). **Input:** ```jsonc { "query": "anything about my previous role at the old company", "purge": false } ``` **Output:** `deleted`, the number of memories soft-deleted. `trace_id` is empty on this path. **Underlying REST endpoint:** `POST /api/v1/{context_id}/forget` --- ## `upload` Upload a document (base64 bytes) into the knowledge layer. Processing is asynchronous. **Requires:** `memory:write`. **Input:** ```jsonc { "bytes_base64": "", // required "title": "Returns Policy", "source": "returns.pdf", "mime_type": "application/pdf", "filename": "returns.pdf", "scopes": [["org/acme/team/eng"]], // optional - or `scope` alias "labels": ["team=eng"] } ``` **Output:** ```jsonc { "id": "doc:01hx9…", "status": "queued", "content_hash": "blake3:…", "deduplicated": false, "version": 1 } ``` Poll status with **`inspect`** (`document:`) or REST `GET .../documents/{id}`. **Underlying REST endpoint:** `POST /api/v1/{context_id}/documents` --- ## `inspect` Fetch a typed record by reference for explainability. **Requires:** `memory:read`. **Input:** ```jsonc { "ref": "entity:person/alice" // entity:/ | trace: | document: } ``` **Output:** Entity (with attributes / supersession), flat trace record, or document metadata - matching the underlying GET. **Underlying REST endpoints:** - `GET /api/v1/{context_id}/entities/{type}/{name}` - `GET /api/v1/{context_id}/traces/{id}` - `GET /api/v1/{context_id}/documents/{id}` --- ## Error handling Operation failures return an **`isError: true`** tool result (HTTP transport stays **200**), not a JSON-RPC protocol error. The result includes the same HTTP status the REST API would return: ```jsonc { "isError": true, "structuredContent": { "error": { "status": 404, "message": "Document not found" } } } ``` | Status | Meaning | Agent action | | --- | --- | --- | | `401` | Missing or invalid key | Fix credentials; missing Context is masked as 401 (not 404) | | `403` | Grant denied | Narrow scope or request access | | `404` | Resource not found | Adjust query or id | | `429` | Rate limit or `enforcement_blocked` | Back off and retry | | `500` | Server fault | Retry; details are redacted | JSON-RPC **`error`** responses are reserved for **protocol faults** only - malformed params (`-32602`), unknown tool (`-32601`), oversized `k`, and similar. Denied operations emit the same **`authz.denied`** audit events and error metrics as REST. --- ## Tool ACL summary | Tool | Typical grant | |---|---| | `remember` | `memory:write` | | `recall` | `memory:read` | | `context` | `memory:read` | | `reflect` (`persist: false`) | `memory:read` | | `reflect` (`persist: true`) | `memory:read` + write region | | `forget` | `memory:forget` | | `upload` | `memory:write` | | `inspect` | `memory:read` | ## Streaming Most tools are synchronous - the full response is returned in a single payload. `upload` returns immediately with `queued` status; ingestion continues on the worker. `reflect` with `persist: true` on a large corpus may take several seconds (no MCP progress notifications today).