recall

Your Claude Code memory, in every agent. recall reads and writes the memory Claude Code keeps for each project, so Codex, Cursor or an agent you build yourself can use the same memories, over MCP or from code. Zero dependencies, no vectors, no database: the memories stay plain Markdown files.

Claude Code now remembers things on its own: your corrections, how you like to work, project decisions, where things live. It saves them as Markdown files with [[wiki-links]] under ~/.claude/projects/<project>/memory/. That is a good format, but only Claude Code reads it. Switch to another agent and it starts from zero.

recall makes that memory portable. It finds the same directory Claude Code uses (same rules, down to its folder-naming hash), reads every file Claude Code writes, and writes files Claude Code reads back.

Quick start: share it over MCP

Point any MCP client at recall and the project's repository. --read-only lets the other agent read what Claude Code remembers without changing it; leave it off to let both write.

Codex

codex mcp add recall -- npx -y @fainir/recall --claude-code /path/to/your/repo --read-only

Cursor (.cursor/mcp.json in the project, or ~/.cursor/mcp.json)

{
  "mcpServers": {
    "recall": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@fainir/recall", "--claude-code", "/path/to/your/repo", "--read-only"]
    }
  }
}

Gemini CLI: gemini extensions install https://github.com/fainir/recall (it serves the memory of the folder you start Gemini in).

Any other MCP client: run npx -y @fainir/recall with the same arguments as a stdio server. Needs Node.js 18+.

The agent then has these tools. Here is a real run against a repository with three memories Claude Code wrote:

> memory_list
3 memories in /Users/you/.claude/projects/-Users-you-code-acme-api/memory:
- feedback_testing (feedback) - run the integration suite before saying a change works
- reference_ci (reference) - where CI runs and how to rerun a flaky job
- user_role (user) - backend lead, prefers short answers with the command to run

> memory_recall {"requires": ["feedback_testing"]}
## feedback_testing (feedback) - run the integration suite before saying a change works
Run `npm run test:int` (needs the local Postgres from docker compose), not just `npm test`.
**Why:** the unit tests mock the database; two "done" changes broke prod queries.
**How to apply:** any change under src/db/ or src/queries/. CI details in [[reference_ci]].

## reference_ci (reference) - where CI runs and how to rerun a flaky job
CI is GitHub Actions, workflow `ci.yml`. Rerun a flaky job with `gh run rerun --failed`.

tokens_used: 180

memory_recall followed the [[reference_ci]] link on its own. In testing, a Codex agent given only this server answered a question by calling memory_list and then memory_get.

Tool	What it does
memory_list	Every memory with its type and one-line description. Optional type filter.
memory_get	One memory in full. feedback-testing and feedback_testing both work.
memory_recall	The named memories plus the ones they [[link]] to, trimmed to a token budget. Unknown names come back as missing.
memory_due	Memories old enough to re-confirm before relying on them.
memory_write	Save or replace a memory in Claude Code's format, and add it to the MEMORY.md index. Not offered with --read-only.
memory_delete	Delete a memory and its index line. Not offered with --read-only.

recall-mcp options: --claude-code [repo] (the default, using the current directory), --dir <dir> for any other memory directory (add --format claude-code to write Claude Code's file format there), and --read-only. RECALL_DIR and RECALL_READ_ONLY=1 work as environment variables.

Use it from code

npm install @fainir/recall

import { Recall } from '@fainir/recall';

const memory = Recall.forClaudeCode('/Users/you/code/acme-api');
console.log(memory.dir);

for (const m of await memory.list()) console.log(m.name, '|', m.type, '|', m.description);

const slice = await memory.recall({ requires: ['feedback-testing'] });
console.log(slice.memories.map(m => m.name), slice.tokens_used);

// Save something Claude Code will see in its next session.
await memory.write({
  name: 'project_freeze',
  type: 'project',
  description: 'merge freeze until the 3rd for the mobile release',
  body: 'No merges to main until Nov 3. **Why:** mobile release cut.',
});

Output (before the write):

/Users/you/.claude/projects/-Users-you-code-acme-api/memory
feedback_testing | feedback | run the integration suite before saying a change works
reference_ci | reference | where CI runs and how to rerun a flaky job
user_role | user | backend lead, prefers short answers with the command to run
[ 'feedback_testing', 'reference_ci' ] 180

write() produces the file Claude Code itself writes:

---
name: project_freeze
description: merge freeze until the 3rd for the mobile release
metadata:
  node_type: memory
  type: project
  modified: 2026-10-27T17:51:14.486Z
---

No merges to main until Nov 3. **Why:** mobile release cut.

and adds a line to MEMORY.md, the index Claude Code loads at the start of every session, leaving the existing lines alone:

- [Testing](feedback_testing.md) - run the integration suite before saying a change works
- [CI](reference_ci.md) - where CI runs and how to rerun a flaky job
- [Role](user_role.md) - backend lead, prefers short answers
- [project_freeze](project_freeze.md) - merge freeze until the 3rd for the mobile release

Rewriting an existing memory keeps the fields recall does not own, such as Claude Code's originSessionId.

Where Claude Code keeps the memory

claudeCodeMemoryDir(repo) returns the directory, using Claude Code's rules:

• The project is the git repository root, so subfolders share one memory, and a worktree uses the memory of its main checkout. Outside git, it is the folder itself.
• The folder name is the project path with every character other than a-z, A-Z and 0-9 turned into -. Paths longer than 200 characters are cut and given the same hash suffix Claude Code adds.
• CLAUDE_CONFIG_DIR replaces ~/.claude. CLAUDE_CODE_PROJECT_DIR_NAME (next to CLAUDE_CONFIG_DIR) names the folder. An autoMemoryDirectory setting in the project's .claude/settings.local.json or your ~/.claude/settings.json wins over both.

This was checked against a real ~/.claude/projects folder: the rule reproduced 40 of 41 project folder names, including all three long, hashed ones (the other one belonged to a session that was moved after it started). recall listed and parsed all 194 memory files in it.

API

The package exports the Recall class and claudeCodeMemoryDir(). Every method is async.

Call	Arguments	Returns
Recall.forClaudeCode(repo)	path inside a project; default the current directory	a Recall on Claude Code's memory directory, in claude-code format
new Recall({ dir, format })	dir: memory directory. format: recall (default) or claude-code	a Recall instance
write({ name, type, description, decay_class, body, extra })	name: slug (required). type: user | feedback | project | reference (required). description: one line, default ''. decay_class: permanent | slow | fast | ephemeral, default slow. body: Markdown, default ''. extra: more frontmatter fields	{ name, type, decay_class, updated_at }. Creates or replaces <dir>/<name>.md; in claude-code format also updates MEMORY.md
get(name)	a memory name; - and _ are interchangeable, and a frontmatter name also matches	a Memory, or null
list({ type })	optional type filter	Memory[], sorted by name
recall({ requires, link_depth, budget_tokens })	requires: names. link_depth: [[link]] hops to follow, default 1. budget_tokens: default 2000	{ memories: Memory[], tokens_used, missing: string[], dropped: string[] }
dueForReview(asOf)	Date or date string, default now	Memory[] past their review window, each with age_days
delete(name)	exact memory name	true if a file was removed, false if it did not exist
claudeCodeMemoryDir(repo)	path inside a project	the directory path (it may not exist yet)

A Memory is { name, type, description, updated_at, frontmatter, body, links }. type and updated_at come from the flat fields or from Claude Code's metadata: block, whichever the file has. links lists the memory names the body links to, deduplicated and in order.

write, get and delete throw on an unsafe name; write also throws on an invalid type or decay_class, a non-string body or description, or a reserved extra key. recall never throws on unknown names; it lists them in missing.

Memory file format

One file per memory, <name>.md, with frontmatter and a Markdown body. recall reads both shapes, whichever format you pick for writing.

Claude Code's format (format: 'claude-code'; older Claude Code files have a flat type: instead of metadata:, and recall reads those too):

---
name: feedback_testing
description: run the integration suite before saying a change works
metadata:
  node_type: memory
  type: feedback
  modified: 2026-10-20T09:12:44.120Z
---

Run `npm run test:int` first. CI details in [[reference_ci]].

recall's own format (format: 'recall', the default, which earlier versions wrote):

---
name: top-customers
type: project
description: paying customers above $10K ACV
decay_class: slow
updated_at: 2026-05-25
---
- Northwind - $48K ACV (see [[northwind-account]])

• Names: a-z, 0-9, - and _, starting with a letter or digit, up to 128 characters. That keeps file names safe on every OS.
• Links: [[name]], [[name|label]] and [[name#heading]] all point at name. They are case-insensitive, - and _ match each other, and links inside code are ignored.
• Frontmatter is a small YAML subset: key: value lines, quoted strings, # comments, and one level of nesting (Claude Code's metadata: block).
• Other files (MEMORY.md, README.md, My Notes.md, folders) are not memories and are skipped.

How recall() picks the slice

1. Every name in requires. No embeddings, no fuzzy search.
2. Every memory reachable through [[wiki-links]] from those, breadth-first, up to link_depth hops.
3. If the total is over budget_tokens (estimated as characters / 4), memories are dropped until it fits: the deepest links first, so the required memories go last; at the same depth, ephemeral first, then fast, slow, permanent. Their names come back in dropped, and the rest keep their walk order.

For an agent that knows what it needs, explicit keys beat similarity search: asking for "voice" should return the voice memory, not four memories that mention the word.

Decay

decay_class says how often a memory should be re-confirmed. dueForReview() returns the ones whose last update (Claude Code's modified, or recall's updated_at) is older than their window. Nothing is deleted automatically.

decay_class	Due for review after
permanent	730 days
slow (default)	90 days
fast	30 days
ephemeral	7 days

Limitations

• Privacy: serving Claude Code's memory to another agent sends those memories to that agent's model provider. Use --read-only and point it only at projects you are happy to share.
• Claude Code reads MEMORY.md at the start of a session, so a memory another agent writes shows up in Claude Code's next session, not the current one. Claude Code loads only the first 200 lines (or 25 KB) of that index.
• Writes are atomic (temp file + rename), but the last writer wins. There is no locking between recall and a running Claude Code session.
• Auto memory is per machine. recall does not sync it anywhere.
• The token count is an estimate (characters / 4), not a tokenizer.

Development

No dependencies, so there is nothing to install:

git clone https://github.com/fainir/recall.git
cd recall
npm test                 # node:test, no network
node examples/basic.js   # the library on a plain recall directory

License

MIT

---

More open source by Nir Feinstein

• agent-harness - make Claude Code or Codex act as a product team, a CEO or a portfolio manager
• frugal - your agent finds the cheaper honest way before it pays for anything
• flipback - an automatic /goal for every turn: sends Claude Code back when it stops early
• most-capable-agent-system-prompt - the copy-paste prompt for a self-improving agent system
• All projects