Coding agent with project memory

This guide shows how to build a coding assistant – similar to an in-editor AI agent – that uses Spectron via MCP to remember coding conventions, past decisions, team preferences, and ongoing work across sessions.

What the agent remembers

A coding agent working on a repository accumulates a surprising amount of context that is useful across sessions:

• •

  **Coding conventions**: "We use `Result` not exceptions in this codebase"

• •

  Past decisions: "We chose Tanstack Query over SWR because of the devtools"

• •

  Team preferences: "Alice prefers functional components; Bob owns the auth module"

• •

  Ongoing work: "The migration to the analytics database is 80% complete, blocked on the auth service"

• •

  Documentation: Product specs, architecture decisions, API contracts

Without persistent memory, every new session starts from scratch. With Spectron, the agent can recall relevant context at the start of each conversation and update its memory as new decisions are made.

Installing the MCP server

Spectron provides an MCP server that exposes memory operations as tools. Install it into your editor using the MCP installer:

# Install for Cursor
npx install-mcp spectron --client cursor --context my-project

# With repository-scoped memory
npx install-mcp spectron \
    --client cursor \
    --context my-project \
    --scope org/acme/project/my-repo

This registers the Spectron MCP server in your editor's MCP configuration. You will be prompted for your Spectron API key during installation.

Manual configuration

If you prefer to configure the MCP server manually, add it to your editor's MCP config:

{
  "mcpServers": {
    "spectron": {
      "command": "npx",
      "args": ["@spectron/mcp"],
      "env": {
        "SPECTRON_API_KEY": "sk-…",
        "SPECTRON_SCOPE": "{\"project\": \"my-repo\"}",
        "SPECTRON_SERVER": "https://spectron.acme.com"
      }
    }
  }
}

Scope per repository

Scoping memory to a repository ensures that conventions from one project do not bleed into another. Use the project dimension in the scope:

["org/acme/project/my-repo"]

With this scope, all memory stored during sessions for my-repo is isolated from memory for other projects. Team-wide knowledge (e.g. company-wide conventions) lives at the org scope and is readable by all project-scoped sessions.

Session start: loading context

At the beginning of each coding session, the agent should call context to load relevant context before it responds to the first user message:

Tool: context
Arguments: {
  "query": "project conventions and active work for my-repo",
  "lens": ["org/acme/project/my-repo"]
}

This returns a markdown context block of what Spectron knows about this project: ongoing work, decisions, preferences, and conventions. The agent includes this in its system context before the user's first message.

Storing decisions

When a significant decision is made during a session, use remember to persist it:

Tool: remember
Arguments: {
  "text": "We decided to use Zod for runtime validation because it integrates with our existing TypeScript types and provides better error messages than Yup.",
  "scope": ["org/acme/project/my-repo"]
}

For team preferences and conventions, the agent can store these automatically when it identifies a standing directive in the conversation:

User: \
  "Always use named exports in this project, never default exports."

Tool: remember
Arguments: {
  "text": "Always use named exports, never default exports.",
  "scope": ["org/acme/project/my-repo"]
}

Documentation lookup with recall

When the user asks a question that might be answered by project documentation ingested into authoritative knowledge, use recall:

Tool: recall
Arguments: {
  "query": "What is the authentication flow for the API?",
  "lens": ["org/acme/project/my-repo"],
  "k": 3
}

This searches the authoritative knowledge base for relevant passages from ingested documents (architecture decision records, API specs, README files).

To ingest project documentation into authoritative knowledge:

spectronctl contexts ingest ctx_acme ./docs/architecture.md \
    --label "Architecture Decision Records" \
    --tags "architecture,adr" \
    --api-key mgmt-…

spectronctl contexts ingest ctx_acme ./docs/api-spec.yaml \
    --label "API Specification" \
    --tags "api,spec" \
    --api-key mgmt-…

Full session flow

Here is what a typical coding session looks like with Spectron MCP integrated:

Session start

Tool: context
→ Returns: ongoing work, conventions, decisions, preferences

Tool: recall (query = user's first message)
→ Returns: relevant documentation passages

During the session

User: "How should I handle errors in the new payment module?"

Tool: recall
Arguments: { "query": "error handling patterns", "lens": ["org/acme/project/my-repo"] }
→ Returns: "Use Result<T, Error> not exceptions in this codebase"

Agent: "Based on this project's conventions, you should use Result<T, Error>…"

When a decision is made

User: "Let's use Stripe's webhook library directly rather than wrapping it."

Tool: remember
Arguments: {
  "text": "Payment module uses Stripe webhook library directly, no wrapper.",
  "scope": ["org/acme/project/my-repo"]
}

Session end (optional)

Tool: reflect
Arguments: {
  "query": "What was decided and what needs follow-up?",
  "persist": true
}
→ Synthesises and stores a session summary

Multi-developer teams

When multiple developers work on the same project, they share memory at the project scope. Each developer's session contributes to and reads from the same experiential memory.

For individual preferences, use the user dimension alongside project:

["org/acme/project/my-repo/user/alice"]

Alice's personal preferences (code style, preferred approaches) live at the user scope. Project-wide conventions live at the project scope. Both are accessible when Alice's session is active.