Memory for Claude Code

Claude Code starts every session from zero. CLAUDE.md holds what you write down; Recall learns the rest from what you say (your preferences, the decisions you made, the people and projects you mention) and brings it back in the next session, with the date and the message each fact came from.

Install

In Claude Code (it needs plugin support and Node 18 or later):

/plugin marketplace add recallmem/recall-plugin
/plugin install recall-memory@recall

Claude Code asks for your Recall API key when you enable the plugin and keeps it in your system's credential store. Create one at app.recallmem.dev; new accounts get $5 of credit. Restart Claude Code, then run /recall-memory:status to check the connection.

What it remembers

• You: "I use pnpm, not npm", "explain diffs before you apply them", your time zone, the stack you prefer.
• Decisions: "we moved staging to a new database in October", "the API stays on v3 until the migration lands".
• People and projects: who reviews which part of the code, what the side project is called, when the launch is.

When something changes, the old fact is closed with the date it stopped being true, not overwritten. Ask "what do you know about me?" in a new session, and every answer points to the message it came from. You can read, correct or delete any fact on the dashboard.

How it works

• When a session starts, resumes or is compacted, the plugin adds what Recall knows about you to Claude's context (about 1,500 tokens by default).
• It records each prompt and Claude's final reply. Recall learns only from what you say; Claude's replies are context, never a source of facts about you.
• It adds eight MCP tools, so Claude can look something up, remember a fact you ask it to, or forget one.
• Tool calls, command output and file contents are never sent. Your code stays on your machine.
• If Recall cannot be reached, Claude carries on. Messages wait on disk, in order, and are sent when it is back; the next session tells you if any were lost.

Recall and CLAUDE.md

Use both. CLAUDE.md is the instructions you write and keep in the repository. Recall is what Claude learns from working with you, across every project, without you writing it down, and it keeps track of when each thing was true.

Codex too

The same repository has a Codex plugin with the same memory, so what you told Claude Code is there when you use Codex:

codex plugin marketplace add recallmem/recall-plugin
codex plugin add recall-memory@recall

Then run /hooks in Codex and trust the three Recall hooks. Other setups are in the docs.

What it costs

Recall learns from each message once, at $1.75 per million tokens. A busy day of 100 prompts and 100 replies of about 300 tokens comes to about $0.10. Adding memory to each session is included. There is no subscription: you top up credit, and new accounts start with $5.

Questions

Is my code sent to Recall? No. Only your prompts and Claude's final replies. Switch recording off for a project with RECALL_RECORD=false, or all hooks with RECALL_ENABLED=false.

Can I run it myself? Yes. The server runs with Docker; point the plugin at your own URL (self-hosting).

Where is my data? In the EU (Frankfurt). It is never used to train models, and you can erase any fact or a whole person.