v0.4.1 (P0-2) — Anthropic CMA-Memory compat shim.
Anthropic shipped Context-Managed Agent (CMA) Memory in public
beta on 2026-04-23 with anthropic-sdk-python 0.97.0 +
claude-agent-sdk 0.1.68. The data model is a Markdown filesystem
tree at <root>/.memory/ with a sibling audit.jsonl log. The
beta is the most direct competitive surface mnemo has seen all
month — operators who pick CMA for Anthropic's branding can keep
using the SDK while mnemo provides:
- The Markdown tree on disk (read-through / write-through /
mirror modes — see [
SyncMode]). - A bridged audit log: every CMA write produces exactly one
mnemoAuditEventwhoseprev_hashchains into the engine's HMAC ledger. - A one-shot importer that walks an existing CMA tree, ingests every Markdown file as a Mnemo memory, and stamps the bridge head into the audit log.
- An export back to a byte-identical CMA tree so users can leave mnemo cleanly.
Tested against claude-agent-sdk-python==0.1.68 (pinned dev-dep
in the repo's Python SDK + a "tested-against" badge in the
README — drift-watcher CI is tracked under issue #cma-drift).