Memory & MCP | GoShipped Docs

Long-term memory recalls earlier facts with pgvector. MCP is optional and off by default. Neither is required to chat.

Two advanced layers, both optional in practice: Memory is on by default and needs OpenAI embeddings. MCP is off until you opt in.

You can leave MCP alone until you need it. Memory you only turn off if you do not want recall — or if you do not have an OpenAI key yet.

Memory

Let your product remember.

Memory lets the AI reuse relevant information from earlier interactions instead of treating every conversation as new. It is not the chat transcript. Transcripts already persist as messages. Memory is a smaller, searched set of facts about the user.

After a turn     →  extract facts → embed → agent_memory
Next question    →  retrieve similar rows → inject into the prompt

What you already have

Piece	Detail
Store	`agent_memory` (pgvector 1536 + HNSW)
Search	`match_agent_memory`
Service	`apps/api/app/services/memory_service.py`
Tool	`search_memory` (same registry as other tools)
User toggle	Settings → AI → `memory_enabled`
Product flag	`features.memory` (default `true`)

When it is retrieved: every chat turn, retrieve_relevant_memories runs before the model call.

When it is written: after the turn, an extraction prompt (EXTRACTION_SYSTEM_PROMPT in memory_service.py) proposes facts; they are saved if they are not near-duplicates.

Effective on/off: product flag and plan memory_enabled and the user toggle.

OpenAI required

Embeddings (and the extraction model default) go through OpenAI — the same constraint as RAG. Doctor requires OPENAI_API_KEY while features.memory is true.

What a builder changes

Goal	Where
Disable for the product	`features.memory: false`
Change who gets memory	`apps/api/app/plans/config.py`
Extraction behavior	`EXTRACTION_SYSTEM_PROMPT` in `memory_service.py`
Match count / similarity	`MEMORY_MATCH_COUNT`, `MEMORY_MATCH_THRESHOLD`, `MEMORY_DUPLICATE_THRESHOLD` in `apps/api/.env`

You usually do not replace the table. You tune prompts and whether the feature exists.

MCP

Connect external capabilities with MCP.

MCP (Model Context Protocol) lets the AI connect to compatible external tools and services through a standardized interface. In GoShipped it is an extension layer. Native chat, web research, URL reading, files, memory, and Deep Research all work with features.mcp set to false.

This is optional and advanced. Skip it for v1 unless you already have a server to attach.

Default: off

"features": { "mcp": false }

Both shipped servers in apps/api/config/mcp_servers.yaml are enabled: false (website_reader, filesystem).

When the feature is off, the API never contacts MCP. Doctor skips the whole MCP group. Credentials are not required.

Turn it on

Set features.mcp to true
Edit apps/api/config/mcp_servers.yaml (or point MCP_CONFIG_PATH at another file)
Set enabled: true only on servers you actually run
Put secrets in apps/api/.env — reference them as ${MCP_WEBSITE_READER_TOKEN}, never in the YAML

mcp_servers:
  website_reader:
    enabled: false
    transport: http
    url: https://website-reader-mcp.vercel.app/mcp/
    headers:
      Authorization: "Bearer ${MCP_WEBSITE_READER_TOKEN}"

Plans also gate MCP (mcp_enabled: Free false, Pro true in the shipped catalog). A Free user will not see MCP tools even if the product flag is on.

Doctor and credentials

Verified in the current doctor:

Situation	Credentials required?
`features.mcp` is `false`	No
MCP on, server `enabled: false`	No — that server’s `${ENV}` refs are ignored
MCP on, server `enabled: true`	Yes — those env names must exist (local file or Vercel Production names)

Production Doctor uses the same rule: only enabled servers, and only when the feature is on.

Native tools stay in apps/api/app/ai/tools/. Prefer a Python tool when the logic is yours. Use MCP when the capability already lives in another process. See Tools.

Tools

Native Python tools — no MCP required.

Configuration

Feature flags and “only pay for what you enable”.

🧠 Memory & MCP