Magnemo vs OpenClaw memory

Your memory lives on your computer, in plain files you can open.

IN THE TREE: test_the_network_line_passes_on_the_package_and_fails_on_a_planted_import · test_free_tier_stdlib_only

"plain Markdown files in your agent's workspace (default ~/.openclaw/workspace)." [S1]; "plain files and one SQLite index" [S2]

The only write verb. Lands in _staging/ as a memory candidate; never touches canon.

IN THE TREE: test_lists_exactly_four_tools · test_promotion_is_not_a_tool · test_the_author_is_the_seat_and_a_keyholder_claim_is_refused

Daily notes: "Agent during work; memory flush; transcript capture". Curated core: "Dreaming consolidation; direct user request". [S2]

Agent writes NEVER reach canonical stores directly — staging only, always.

IN THE TREE: test_staging_is_the_only_note_write_path · test_retrieve_excludes_staged

not stated in their docs (read 2026-10-02). Docs describe machine "gated consolidation" and "The write path is the security boundary", not a person's approval. [S2]

Every note is markdown with provenance frontmatter (author, written, source, status, reviewed_by, supersedes, strength).

IN THE TREE: test_stage_lands_in_staging_with_provenance_and_salience · test_paperless_and_malformed_notes_refused · governance.promote

SQLite provenance columns: origin class (closed set owner, agent, untrusted, system), session kind, "Observed timestamp and supersession key" for lineage. No author field named. [S2]

Anything that speaks MCP over stdio mounts the same way.

IN THE TREE: test_connect_and_list_over_stdio

Markdown, "inspectable and editable with a text editor" [S2]; import "from Codex, Claude Code, and Hermes" via Control UI [S1]. Export command not stated.

Python 3.11+, zero dependencies, Apache-2.0.

IN THE TREE: LICENSE

MIT (README: "MIT © OpenClaw Foundation") [S3]