v0.2.2

Fixes

• Summarizer no longer records hook-error output. summarize_one() accepted whatever headless claude -p printed on a 0 exit code; a misbehaving hook in the user's environment (e.g. claude-mem's UserPromptSubmit failing) could emit a block notice to stdout that got saved verbatim as a file's summary. Added _is_valid_summary() to reject hook/permission notices before recording (+test).
• Synced __version__ in __init__.py (was stuck at 0.1.0).

Published to PyPI via Trusted Publishing on release.

Cerebro 0.2.1

Fix: SessionEnd hook crash under Python 3.9

The cerebro-session-end.py plugin hook crashed at import when run by the system /usr/bin/python3 (3.9.6) because it used PEP 604 union syntax (str | None) without from __future__ import annotations:

TypeError: unsupported operand type(s) for |: 'type' and 'NoneType'

This silently disabled the auto-record multi-brain sweep at session end. Fixed by adding the future import (matches the project convention in CLAUDE.md).

Patch release: no changes to the src/cerebro MCP server — this ships via the plugin/marketplace channel. Version bumped in lockstep across pyproject.toml, marketplace.json and plugin.json.

Cerebro 0.2.0

What's new in 0.2.0

Self-sustaining brain

• Auto-record (SessionEnd hook): re-summarizes files whose summary went stale during the session, so the next session reuses fresh traces instead of re-reading. Opt-in via CEREBRO_AUTORECORD=N. New cerebro summarize --stale.
• Polyrepo auto-detect: resolves the brain by walking up to the nearest .cerebro/, so a monorepo subfolder uses the shared root brain automatically — no per-folder CEREBRO_ROOT needed.

Front↔back visibility

• cerebro_endpoints: index of the backend HTTP endpoints (NestJS @Controller / @Get / @Post / ...) — answers "where is POST /carts/lines handled?" without grepping decorators.

Token-saving + anti-hallucination (bundled in the plugin)

• cerebro-first (PreToolUse): nudges toward cerebro_search before raw grep/find (once per session).
• verify-edit (PostToolUse): deterministic syntax / undefined-symbol / type check after edits (Python, JS, Dart).
• The plugin now ships the MCP server + these hooks + 5 cerebro-first subagents + a skill.
• Secrets hygiene: .env* files are never indexed.

Install

claude mcp add cerebro -- uvx cerebro-code-memory

or the full Claude Code plugin:

/plugin marketplace add marcodavidd020/cerebro-code-memory
/plugin install cerebro@cerebro

Cerebro 0.1.0

Persistent code-knowledge memory across AI chat sessions. An MCP server that caches a codebase's structure and your understanding of it in a local SQLite "brain", so new sessions query it instead of re-reading folders — token-cheap and 100% local.

Install

claude mcp add cerebro -- uvx cerebro-code-memory

Or the full Claude Code plugin (MCP server + session hooks + cerebro-first subagents):

/plugin marketplace add marcodavidd020/cerebro-mcp
/plugin install cerebro@cerebro

Highlights

• Structural map — tree-sitter symbols + import graph with PageRank centrality; resolves tsconfig/jsconfig aliases. Languages: Python, JS/TS (JSX/TSX), Dart/Flutter.
• Cached summaries — 1–3 sentence English summaries per file, reused across sessions; hash-based freshness flags stale entries.
• Decision log — cerebro_note / cerebro_recall capture the why behind the code.
• Optional local semantic search — model2vec embeddings (no torch, no API key, nothing leaves the machine); hybrid semantic + keyword to the exact path:line.
• Architecture insights — impact (blast radius), cycles, orphans, dead symbols, and a name-resolved call graph.
• Git-aware sync across nested repos; HTML dependency graph + Obsidian export; living-docs audit.

Privacy: no torch, no API keys, nothing leaves your machine. Requires Python ≥ 3.10.

📦 PyPI: https://pypi.org/project/cerebro-code-memory/