Repository navigation
v2.1.1
The summary is stored whole again
2.1.0 cut a session summary at 3,000 characters. The reason it was there —
boot re-reading three summaries in full — had been solved in the same release
by reading only their first 240 characters, so the cut only ever lost the tail
of what the caller wrote. It is gone. The response now carries
summary_chars and, past 3,000, a note with the fact that still matters:
session logs are not searched by recall, so what lives only in a summary is
invisible to it. The facts belong in crbro_learn.
Initiative, in the two places Claude Desktop reads
A field test in Claude Desktop passed everything except initiative: in a new
conversation the model did not boot the memory on its own, and asked whether
semantic recall was on it answered that it had no way to know — although
view=status reports exactly that. Claude Desktop does not read the
instructions a server returns at initialize
(anthropics/claude-code#43749),
so the two levers it does honour carry the message now: crbro_recall's
description says to call it before answering anything about the user, their
projects, preferences or past work, and crbro_inspect's says view=status
answers questions about CRBRO itself. instructions are set as well, for the
clients that read them — Claude Code among them.