Skip to content

v2.1.0 — a memory you query, not one you dump

Choose a tag to compare

@Octonove Octonove released this 07 Sep 00:18
· 142 commits to master since this release

A memory you query, not one you dump

2.0.3 put a ceiling on every read. 2.1 removes the reason the ceiling was
being hit: reads that returned everything when the caller needed one thing.
Every figure below is the text a Claude Code session actually reads
(content[0].text; that client does not pass structuredContent to the
model), measured on the same 1,145-neuron brain, 2.0.3 → 2.1.0.

crbro_inspect view=neuron is an index by default. Header, counts,
connections, and every entry — map first, then errors, debts, preferences,
patterns, decisions and the long tail of facts — as an id, a kind, a date and a
160-character preview, 25 per page with limit/offset. Retired entries are
hidden and counted in entries_pagination.hidden_retired; with
include_superseded=true they come back with revised and retired_note.
entries=[ids or exact text] returns just those in full and names what it
could not find; detail=full is the old whole-neuron read, still budgeted.
The 307-fact neuron: 5,904 tokens shortened in 2.0.3 (66,952 whole in 2.0.2)
→ 1,887 for the complete index.

crbro_recall hands you the handle. Every result and every also_matched
line carries entry_id — stamped on each chunk at index time, so a decision
whose chunk is "text — rationale" still resolves to the decision's own id. The
default is five ranked results, down from ten (the metric is recall@3);
matched_neurons and has_more say how many neurons matched before the cut,
so five never reads as "only five". A hit longer than 1,200 characters comes
back as its opening with content_truncated and content_chars; the
also_matched lines are 300-character previews with their own ids. The
~90-token hint that repeated on every call is one line. Ten results:
6,242 → 3,941 tokens; five, the default: about 2,000.

crbro_boot carries headlines. The last three sessions come as their
opening 240 characters with summary_truncated and summary_chars beside
them; hot_topics are ten rows with the day, not twenty with the
millisecond; active_context no longer repeats open_items and
recently_closed, which boot already serves at the top level, capped at 12
and 8 with _total when there are more. The protocol block,
memory_discipline and retired_tools are untouched. 4,989 → 2,758 tokens.

crbro_inspect view=sessions honours offset and caps each summary at
3,000 characters, declared
— except limit=1, which returns one log whole:
that is the door boot points at.

crbro_consolidate caps the summary at 3,000 characters, after redacting.
The card asks for a line; the field was taking whole reports, median 6,829
characters on the reference brain, re-read at every later boot. Credentials
are redacted before the cut so none is left half-written on disk; the
response returns summary_truncated: {kept_from_this_call, sent} with the
reason, and the parameter now says where the facts belong: in crbro_learn.

Compact JSON on the wire. Responses were pretty-printed; the indentation
was 4–27% of what the model paid, depending on the view (12% of boot, 27% of
global_map). view=global_map: 2,111 → 1,549 tokens.

memory_discipline teaches the habit in one sentence and dropped a
371-character lifecycle paragraph already carried by learn, revise and forget.
The search index is format 6 and rebuilds itself once on the first boot.
Brains on disk are untouched; a client that wants the 2.0 shape passes
detail=full. Eighteen new or adapted tests, 293 in total; the retrieval
benchmark is unchanged at 71% / 77% (79% / 85% with also_matched), now
matching hits by entry_id.