Bring Claude Code’s memory to every supported agent

Claude Code’s auto-memory is useful, but it stays inside Claude Code. Agentcairn v0.25 adds a one-way bridge into the same Markdown vault used by Codex, Cursor, OpenCode, Hermes, and other supported hosts.

The bridge complements Claude’s native memory rather than replacing it: Claude keeps learning as before, while you choose when to copy that durable context into shared cross-agent recall.

Auto-memory was introduced in Claude Code 2.1.59 and must have created files for this repository before there is anything to import. See Claude Code’s auto-memory documentation to inspect or manage it with /memory.

Run the importer from a repository where Claude Code has created auto-memory. Preview is the default and performs no writes:

The report shows what would be added, updated, left unchanged, repaired, or expired. Apply that exact plan only when it looks right:

• Claude auto-memory only. MEMORY.md and its topic Markdown files are eligible.
• Instructions stay out. CLAUDE.md, CLAUDE.local.md, and .claude/rules/ are not memories and are never imported.
• One repository at a time. The current repository is the default. Use --project ../other to target another one.
• Explicit source when needed. Use --source <dir> for a custom or managed Claude memory directory. Ambiguous discovery fails closed instead of guessing.

Imported notes record Claude Code, the project, and the source file as provenance. Recognized credentials are redacted before the plaintext Markdown is written. Model-authored context remains labeled as imported evidence rather than being mistaken for a user instruction.

Re-running the command is idempotent for unchanged files. A changed file creates a new version and supersedes the old one; a source file that disappears expires its imported note. The prior versions remain inspectable instead of being silently overwritten.

Agentcairn stores lifecycle metadata inside .agentcairn/native-memory/ in the vault. That registry also prevents an unchanged source from resurrecting a note you deliberately deleted.

After --apply, the importer normally reconciles the DuckDB cache so supported hosts connected to that vault can recall the new Markdown. If embedding or reconciliation fails, the CLI warns instead of hiding the problem: the Markdown remains durable, and the next successful reconcile makes it available for indexed recall. Codex can then recall a decision Claude learned; Cursor can search it; Obsidian can show you the underlying note and provenance.

Everything stays local with the default FastEmbed configuration. If you opt into a cloud embedder, redacted note chunks and recall queries are sent to that provider. If you enable the Anthropic durability judge for ingest or sweep, candidate redacted session text is sent for judgment.

See supported agents · Inspect memory in Obsidian · How cross-agent memory works