Context Databases vs MCP Layers: Where Memory Belongs

Separate remembered meaning from current system state

Consider an agent preparing a billing reconciliation in a fresh VM. It needs the team's definition of billable revenue, the reconciliation procedure and today's invoice records. The definition and procedure should survive yesterday's session. Invoice amounts should come from the billing system when the task runs.

For this comparison, a context database means the durable store and retrieval service for agent context. An MCP layer means the service exposing or routing MCP operations. These are architectural responsibilities; evaluate which ones a product actually implements.

The distinction avoids two mistakes: treating a search index as a live integration, and assuming that routing tools automatically creates durable memory.

Make the stored record richer than its embedding

A useful billing memory needs a stable identifier, workspace, owner, body and revision history. Add the source of the revenue definition, when it was recorded and whether another decision supersedes it. A similarity score cannot tell the agent whether finance has approved the rule.

Search can start with ordinary text retrieval. PostgreSQL documents stored generated tsvector columns and GIN indexes. This illustrative schema stores text and its search representation together; it is a storage example, with authorization policies still required:

CREATE TABLE agent_context (
  workspace_id uuid NOT NULL,
  path text NOT NULL,
  body text NOT NULL,
  updated_at timestamptz NOT NULL DEFAULT now(),
  search_vector tsvector GENERATED ALWAYS AS
    (to_tsvector('english', body)) STORED,
  PRIMARY KEY (workspace_id, path)
);

CREATE INDEX agent_context_search
  ON agent_context USING GIN (search_vector);

The generated column follows changes to body; updating updated_at remains the application's responsibility. Add revision and provenance records before treating this as a production memory design. If semantic retrieval later helps agents find differently worded revenue rules, retain the original text and identifiers so search representations can be rebuilt.

Ask a context-store vendor how edits, deletion and export affect every search representation. Ask an MCP-layer vendor where those records persist. A tool named remember answers neither question by itself.

Use MCP as the access contract, including for context

The MCP tools specification defines discovery through tools/list and invocation through tools/call. The resources specification also defines URI-addressed context and retrieval through resources/read. MCP therefore supports more than forwarding actions to business applications.

You could expose the revenue definition as a resource and offer a search tool over stored decisions. The backing store remains responsible for durability and record lifecycle. The MCP adapter translates requests into authorized storage operations and returns their results.

For a remote implementation, Claude Code documents this connection syntax. Here the hostname is a placeholder for your service:

claude mcp add --transport http company-context https://context.example.com/mcp

Complete the server's required authentication and check its status through /mcp. Adding the URL configures a connection; it does not create a database, populate memories or arrange automatic recall. When evaluating a combined service, inspect those behaviors independently. Choosing resources, prompts or tools for context covers the interface decision.

Keep skill delivery and prompt loading explicit

Storing the reconciliation procedure does not ensure that an agent loads it. The Agent Skills specification describes a directory containing SKILL.md, with YAML frontmatter and Markdown instructions, plus optional supporting files. A database storing only the Markdown body may lose metadata or relative references needed to reconstruct the skill.

Choose whether the store preserves a complete skill package or offers remotely readable instructions. Then test how each client discovers and activates it. Returning skill text through MCP does not by itself install that directory into a client's native skills location.

Anthropic's context-engineering guidance describes keeping lightweight references and fetching details when needed. Apply that approach here: present a short catalog, retrieve the reconciliation procedure for a billing task, and load the revenue decision when relevant. Putting the whole store into the prompt turns durable storage into a recurring context cost.

A database comparison should therefore include retrieval quality and export fidelity. An MCP-layer comparison should include discovery and client behavior. Neither a large storage allowance nor a successful connection proves the right procedure reaches the model.

Align permissions without confusing memory with authority

The reconciliation agent may read the shared revenue definition and invoices while lacking permission to change either. Model these separately: context reads, memory edits and billing writes. A retrieved procedure saying to issue a refund must not grant the refund capability.

Enforce workspace and personal-record visibility in the context service before returning search results. PostgreSQL row security can restrict rows returned or modified, but table owners normally bypass it, and superusers and roles with BYPASSRLS always do. Test with the actual service role; enabling a policy is insufficient evidence of isolation.

Keep billing credentials with the integration service rather than inside the memory body. Conversely, a gateway's control over billing credentials says nothing about who may read private reconciliation notes. A combined product needs both permission models, even if they share one identity system.

Define freshness and failure behavior across the boundary. Return the recorded date and source with a revenue rule; request current invoice state from billing. If billing is unavailable, report that failure instead of silently substituting a remembered balance. If memory search fails, distinguish unavailable context from an empty result. Colocation can simplify operations, but it also puts retrieval and live access behind a common dependency.

Where SPYN3 combines the two responsibilities

SPYN3 provides one workspace MCP endpoint for connected systems. An admin can enable Memory & skills, off by default, to add personal and shared memories, skills and notes through MCP context tools. These live in its EU database, with Postgres full-text search by default; semantic search needs the workspace's own model key. Context reads need call:tools, memory writes need write:memory, and connected-system writes need call:write_tools. Upstream credentials stay in the workspace with envelope encryption. This combines storage and access while keeping their permissions distinct.

Choose by the responsibility you cannot already satisfy

If your agents already reach billing reliably but repeatedly lose agreed revenue definitions, prioritize durable context: ownership, revision history, recall and export. If definitions already live in a maintained repository but every VM needs separate billing setup, prioritize the integration layer. Choose a combined service when both gaps exist and its individual contracts meet your requirements.

Before committing, run one reconciliation exercise against the proposed architecture:

1. Record a fictional revenue rule, then retrieve it from a fresh agent without copying the previous conversation.
2. Revise the rule and confirm that retrieval identifies the current record and preserves its provenance.
3. Change a fictional invoice upstream and verify that the agent obtains the new amount through live access.
4. Attempt to edit the rule and issue a refund with read-only authorization; require both to fail.
5. Export the rule and procedure, then demonstrate how another service could serve them.

These checks reveal whether you are buying durable context, controlled connectivity or both. Continue with designing an MCP memory server to specify the storage-facing operations before choosing an implementation.