recall: give Codex the memory Claude Code already wrote for your repo (MCP, zero deps)

[fainir]

If you switch between Claude Code and Codex in the same repo, Codex starts from zero: Claude Code's memories (your corrections, how you like to work, where things live) sit in Markdown files under ~/.claude/projects/<project>/memory/ that only Claude Code reads.

I built recall, a zero-dependency MCP server that gives Codex those same memories:

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

Codex then gets memory_list, memory_get, memory_recall (follows [[wiki-links]] between memories, trimmed to a token budget) and memory_due. Drop --read-only to add memory_write and memory_delete, which write in Claude Code's own format so both agents share one memory.

A real run against a repo with three memories Claude Code wrote: memory_recall {"requires": ["feedback_testing"]} returned that memory plus the CI note it links to, in 180 tokens.

No vectors, no database, nothing leaves your machine. It finds the same folder Claude Code uses, down to its folder-naming rule.

Repo (MIT): https://github.com/fainir/recall - also on npm as @fainir/recall and in the official MCP Registry.

Feedback welcome, especially from anyone who runs Codex and Claude Code side by side.

[FramY2]

Really clean design @fainir — keeping the shared memory layer as plain Markdown files under ~/.claude/projects/<project>/memory/ with [[wiki-links]] traversal and zero vector-DB dependencies makes a lot of sense for repos where developers alternate between Claude Code and Codex.

We profiled @fainir/recall (v0.1.3) with ctxwise mcp-xray (npx @framy2/ctxwise mcp-xray --probe-stdio) to measure the per-turn MCP tools/list schema footprint in both modes:

Mode	Tools Exposed	Total Schema Chars	Est. Startup Tokens / Turn	[BLOATED] Tools (>= 250 tokens)
Read-Write (default)	6 (memory_write, memory_recall, memory_due, memory_list, memory_get, memory_delete)	2,935 chars	~735 tokens/turn	1 (memory_write at 1,155 chars / ~289 tokens, 7 params)
--read-only	4 (memory_recall, memory_due, memory_list, memory_get)	1,545 chars	~387 tokens/turn	0

Two takeaways from testing it side by side:

1. --read-only cuts per-turn schema overhead by -47.3% (-348 tokens/turn): Since memory_write carries the heaviest schema description and parameter block (~289 tokens by itself), defaulting the Codex registration in your README to --read-only is a great call — it keeps every tool well below the bloat threshold while still giving Codex full graph traversal via memory_recall.
2. Pre-flight secret redaction before returning m.body across agents: In practice, Claude Code auto-memories occasionally capture raw debugging artifacts (e.g., ://user:pass@ database DSNs, sk-ant-* / sk-proj-* keys, or ghp_* tokens) when a user tells Claude "remember how we fixed the staging connection". Adding a lightweight regex redaction pass inside memory_get / memory_recall before returning m.body (or running npx @framy2/ctxwise shield against the memory directory) is an easy way to ensure credentials written during a Claude session never leak into Codex's prompt context.

[fainir]

Thanks, both points are fair. The README already registers Codex with --read-only.

On secrets: recall now redacts them before a memory leaves the MCP server (fainir/recall#1). Passwords in connection URLs, bearer tokens, private key blocks, Anthropic, OpenAI, GitHub, Slack, AWS, Google and Stripe keys and JWTs come back as [redacted], the reply says how many were hidden, and the files on disk are untouched. --allow-secrets turns it off. It is on main and will be in the next npm release (0.3.0).

[FramY2]

Awesome turnaround @fainir — just read through fainir/recall#1 (0.3.0) and the implementation in bin/recall-mcp.js is super clean.

A few details that stood out:

• Covering all four read surfaces (memory_get, memory_recall, memory_list, memory_due) at the egress boundary while leaving the Markdown files on disk untouched and keeping [[wiki-links]] traversal working before redacting the assembled output.
• Preserving URL structure (postgres://user:[redacted]@host/db) + explicit redaction count footer: Appending [N secret(s) redacted; start recall-mcp with --allow-secrets to include them] is a great UX touch for agents — Codex immediately sees that the credential was intentionally masked rather than assuming the DSN is malformed.
• Runtime-constructed test fixtures (rep(c, n) in test/redact.test.js): Building the synthetic sk-ant-*, sk-proj-*, ghp_*, AKIA*, and JWT tokens dynamically at test time keeps GitHub push protection and static secret scanners completely happy.

Congrats on shipping 0.3.0!