Supermemory for Claude Code

The Supermemory plugin adds an external memory path to Claude Code. It can capture configured conversation context and retrieve useful earlier work, including project decisions and preferences, across sessions.

Claude Code also has native instructions and auto memory. The reason to add a plugin is a specific context-sharing or retrieval requirement that your current setup does not satisfy.

Start with a concrete continuity problem

Suppose a developer records why a function intentionally uses an unusual implementation. A later task should retrieve that rationale before refactoring it. The agent must still inspect the current code and check whether the decision remains applicable.

Keep the decision linked to its source so a later task can check whether it still applies.

Install the plugin

The current plugin README requires Node.js 18 or newer for its hooks and documents these Claude Code commands:

/plugin marketplace add supermemoryai/claude-supermemory
/plugin install supermemory

Set SUPERMEMORY_CC_API_KEY using your supported server or shell secret configuration. Use /supermemory:status to inspect authentication and /supermemory:project-config for repository settings. If migrating an older plugin name, follow the README's migration instructions rather than assuming it updates in place.

The coding plugin is free. Hosted memory operations consume usage under current pricing, and the answering model has its own charges.

Understand capture and recall separately

The plugin documents automatic capture and reasoned recall. Before a turn, the agent is prompted to decide whether earlier context would help; a search is not guaranteed on every message. Capture settings determine what is submitted for future use.

A conversation can contain proposals, rejected alternatives and quoted instructions. Inspect what is retained and verify that a suggestion does not become a confirmed project decision. Enable or adjust selective capture according to the workflow's needs.

When recall fails, inspect processing status, scope and retrieved records to find where the context was lost.

Keep repository and personal scope explicit

The current coding plugins derive a shared repository container identity from the normalized Git remote, with configuration overrides and a path fallback. Personal and project records use scope metadata within that design.

A shared container name does not grant access across accounts. Compatible credentials and authorization are still required. Verify which scope a second client can read before treating a remembered team decision as shared knowledge.

If worktrees need separate context, inspect the documented isolation setting. Test a different repository with the same folder name to ensure the intended boundary is preserved.

How this differs from connecting MCP tools

MCP provides a tool connection. A plugin can also attach behavior to events in the host application, such as conversation capture. Simply connecting a memory server does not guarantee automatic saving, while exposing a tool does not prevent a client from adding its own automation.

Treat the plugin and MCP as separately configured paths. If both are enabled, inspect their accounts, scopes and write behavior so the same transcript is not saved twice unintentionally.

Verify the experience across sessions

Save a fictional project decision and end the session. Start a new session in the intended repository and explicitly search for it. Inspect the returned evidence, then ask a related question.

Change the decision, repeat the search and check the answer. Test removal through the documented lifecycle, including any retained local notes and conversation history. An old fact appearing in an answer may come from a different copy than the record just removed.

The memory lifecycle guide and coding-agent retrieval guide cover those checks in more detail.

Evaluate the plugin on your own work

A public memory benchmark does not establish a coding success rate or a guarantee for every repository. Measure whether the plugin helps with your actual tasks: recovering an earlier rationale, respecting a corrected convention and avoiding an obsolete implementation.

Install using the maintained README when those behaviors are useful. Keep native project instructions, current source inspection and code validation in the workflow, then judge the additional memory by the evidence it contributes.

Frequently asked questions

Does Claude Code have native memory?

Yes. Claude Code documents project instructions and auto memory. The Supermemory plugin adds a separately configured external memory path.

What should you check when the plugin misses a detail?

Check that the detail was captured, processing completed and the current session uses the intended scope. Then inspect whether retrieval returned the record and whether it reached the answer.

Does MCP itself prevent automatic capture?

No. MCP supplies a tool connection; host-specific automation can add capture. Simply connecting a server does not establish that such automation exists.