v2.0.3 - Fix sqlite3 missing in installer exe

Bug Fix

Fixed 'No module named sqlite3' error when initializing a project via the installer exe.

Root cause: PyInstaller only traces static imports. Since db.py is bundled as a data file (not analyzed), its import sqlite3 was invisible to PyInstaller. Fixed by adding --hidden-import sqlite3 --collect-all sqlite3 to force inclusion.

Downloads

• cc-memory-installer.exe — Installer + project initializer (11.7 MB)
• cc-memory-dashboard.exe — Visual memory dashboard (12.0 MB)

v2.0.2 - Fix exe project initialization

Bug Fixes

• Fixed SQLite error when initializing project via exe — The installer now uses the bundled db.py (from inside the exe) instead of requiring the plugin to be installed first. You can now initialize a project in Step 3 without doing Step 1 first.
• Uninstall button now visible — The Uninstall button in the bottom toolbar has been present in code since v2.0.1 but required an exe rebuild to appear. This release includes the rebuilt exe.

Downloads

• cc-memory-installer.exe — One-click installer + project initializer (10.8 MB)
• cc-memory-dashboard.exe — Visual memory management dashboard (12.0 MB)

v2.0.0 — Real-time Observer, FTS5, MCP Server, Web Viewer

cc-memory v2.0.0

Major upgrade matching claude-mem feature set while keeping zero external dependencies (pure Python stdlib).

New Features

• Real-time observer — Stop hook calls Haiku AI after each Claude response to evaluate observations and decide what to save (like claude-mem's observer agent, but batched per-turn)
• PostToolUse capture — Every tool call recorded as observation in SQLite
• FTS5 full-text search — SQLite native ranking with LIKE fallback
• Progressive disclosure — 5-layer token-budget context injection (topics → critical → timeline → handoff → footer)
• SHA-256 dedup — Content hash deduplication replaces string comparison
• Structured session summaries — 6 fields (request/investigated/learned/completed/next_steps/notes)
• Privacy tags — <private> content stripped before storage
• MCP server — 7 tools via JSON-RPC stdio
• Web dashboard — localhost:9377 browser-based SPA
• Mode system — code/research/writing profiles
• File logger — All hooks log to file, never stderr (fixes hook error UI)
• One-click uninstall — GUI button + --uninstall CLI flag

Database

• 8 backward-compatible migrations (auto-applied on first connect)
• Existing memory.db files upgrade transparently
• SQLite PRAGMA optimizations (busy_timeout, mmap, cache)

Three-Layer Save Mechanism

Layer	When	Who Decides
Observer	After each Claude response	Haiku AI (real-time)
PreCompact	Before compaction	Haiku AI (batch)
/save-memories	Manual	Claude (on-demand)

Installation

Download cc-memory-installer.exe and run it. Three steps:

1. Install Plugin → copies files to ~/.claude/hooks/cc-memory/
2. Configure Hooks → registers 5 hooks in settings.json
3. Initialize Project → creates memory/ in your project

Files

• cc-memory-installer.exe — One-click installer (GUI + CLI)
• cc-memory-dashboard.exe — Tkinter visual dashboard

v1.2.0 — Topic-based memory consolidation

What's New

Hierarchical Topic Consolidation

Memories are now organized into a 3-level hierarchy instead of a flat list:

• Level 0: Topic summaries — consolidated via Haiku LLM, always injected at session start
• Level 1: Active memories — top-5 per topic kept, rest archived
• Level 2: Archived — queryable but not injected

New Module: consolidate.py

6-step consolidation pipeline:

1. Garbage cleanup — regex patterns remove Claude meta-text, XML tags, etc.
2. Fuzzy dedup — trigram Jaccard similarity (threshold=0.65) merges near-duplicates
3. Topic assignment — keyword-frequency-based, no LLM needed
4. LLM summarization — Haiku generates per-topic summaries → topics table
5. Importance decay — imp=5 decays to 4 after 30 days, 4→3 after 60 days
6. Archive consolidated — memories captured in summaries get archived (keep top-5)

Improved Context Injection (session_start.py)

• Injects topic summaries (compact, ~250 chars each) instead of raw memory dump
• Adds unmerged critical memories + recent high-value items
• Total injection: ~4.5KB structured vs ~3KB random before

Smarter Extraction (pre_compact.py)

• LLM extraction now includes topic field for each memory
• Auto-consolidation triggers every 5 sessions

New CLI Commands (mem.py)

mem.py --project <path> topics        # View all topic summaries
mem.py --project <path> consolidate   # Run full consolidation
mem.py --project <path> cleanup       # Quick: garbage + dedup + topic assign (no LLM)
mem.py --project <path> add ... --topic <name>  # Add with topic tag

Schema Migration

• memories table gains topic TEXT column (auto-migrated on first connect)
• Migration tracking via _migrations table
• Backward-compatible: existing DBs upgrade transparently

Results on real project (288 memories, 39 sessions)

Metric	Before	After
Active memories	288 (flat)	76 (top-5/topic)
Topic summaries	0	16
Near-duplicates	9+	0
Injection quality	Random dump	Structured by topic

Full Changelog: v1.1.1...v1.2.0

v1.1.1 — Fix non-ASCII path encoding bug

Fix

• Fixed encoding bug with non-ASCII project paths — When the project directory contains non-ASCII characters (e.g., Chinese D:\档案\), json.dumps() with default ensure_ascii=True produced Unicode escape sequences that created invalid surrogate pairs in API request bodies, causing:

  • API Error 400: "invalid high surrogate in string"
  • Empty memory.db (0 memories, 0 sessions saved)
  • Garbled path display in logs

• Added ensure_ascii=False to json.dumps() in both session_start.py and pre_compact.py API request serialization.

Files

• cc-memory-installer.exe — One-click installer
• cc-memory-dashboard.exe — Memory management GUI

v1.0.0: Auto-deploy /save-memories skill and Stop hook on project init

Install with cc-memory-installer.exe. cc-memory-dashboard.exe is a stand alone memory/task manager for this plugin.