MCP hosts now see three tools: MemHub_recall, MemHub_write and
MemHub_diagnostic_trace. MemHub_ask, MemHub_conversations,
MemHub_context, MemHub_memory_write and MemHub_memory_write_status are
removed, with their broker routes (/ask, /conversations, /memory-writes)
and client functions. Broker /context stays: the WorkBuddy injection
hook, recall-cli --context and the Hermes provider still use it.
MemHub_write saves one hand-added memory exactly like the app's
"添加记忆": broker POST /memory-notes authenticates the installation and
calls Core /api/v1/memory-notes with trusted headers, so the note takes
the direct-write lane ahead of history and is reconciled through
add_extracted. It takes {text, request_id?} and returns the note receipt
with the request_id to retry with.
MemHub_recall falls back to the SDK raw layer when no generation model
is configured, or when the memory search is empty and the library holds
no memory yet. The receipt says so with mode "raw" and fallback_reason;
raw hits are conversation excerpts (content_format "raw", source,
session_id). Raw reads keep the /context raw-lane authority: a host
needs om_conversations and its granted, unrevoked sources, so recall's
own grant never widens to original conversations.
Stored grants holding om_ask or om_conversations still load; new
pairings get om_recall, om_conversations and om_diagnostic_trace.
Skill text (SKILL_VERSION 19), connection prompt, docs and tests follow.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
alvinfeiandClaude Opus 5.5
a08f69a747