v1.1.0

What's Changed

• v1.1 by @d2a8k3u in #7

Full Changelog: v1.0.0...v1.1.0

v1.0.0

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

1.0.0 - 2026-03-06

Added

• Cross-encoder reranking — memory_search now reranks results with a TinyBERT cross-encoder (ms-marco-TinyBERT-L-2-v2) for significantly more accurate relevance scoring. Over-fetches 3x candidates, reranks, then slices to requested limit. Graceful fallback if model unavailable.
• Relevance filtering with strictness levels — new strictness parameter on memory_search (low, normal, high) controls minimum topic and relevance thresholds, filtering out noise before results reach the caller
• Topic splitter — automatic decomposition of large multi-topic memories into focused, single-topic units
  • Splits by H2/H3 headers, bold-colon patterns, and numbered bold items
  • Validates preambles, filters short sections
  • Integrated into memory_store, memory_store_batch, and session-end consolidation
• Merge utilities module (merge-utils.ts) — extracted reusable merge logic: tag normalization, safe JSON parsing, intelligent merge updates
• Reorganize command — new node dist/cli.js reorganize subcommand for batch splitting of large memories with relation preservation and accumulative record deduplication
• Memory reorganize skill — guided manual reorganization workflow (5-step process)
• Memory cleanup skill — guided cleanup workflow: deduplicates accumulative records (tech-stack, active-modules), consolidates near-duplicate episodics, removes junk procedurals and low-value patterns
• Tag-based memory lookup — findMemoryByTag() for efficient singleton record retrieval (used by session-end for accumulative records)
• Type-filtered deduplication — findSimilarMemory() now accepts optional type parameter for more precise duplicate detection
• Database transaction helper — transaction() wrapper for atomic multi-step operations
• Auto-split during consolidation — session-start consolidation now splits up to 5 large memories per run
• Reranker model warmup — pre-loads cross-encoder at session start alongside embedding model

Changed

• Search scoring overhaul — FTS/vector blending now uses signal-aware logic: FTS-only matches use full FTS score instead of being diluted by 0.4 weight when no vector match exists. FTS match floor (0.05) prevents common-term IDF collapse from hiding relevant results.
• Session-end singleton records — tech-stack and active-modules are now merged in-place via tag lookup instead of creating duplicates every session. Procedural workflows similarly update-or-create by category.
• Session-end episodic filtering — skips meaningless records (short tasks, interrupted requests with no substance) to reduce noise
• Pattern clustering improvements — consolidation now filters cluster candidates to session-end records only and excludes interrupted sessions
• memory_store now calls topic splitter before storage, with type-filtered dedup and contradiction detection
• memory_store_batch pre-expands splittable items (cap: 100) with within-batch and DB deduplication, plus sibling relations
• Session-start uses relevance thresholds to filter injected context (topic: 0.08, relevance: 0.15)
• Tool descriptions updated to instruct English-only content for consistent cross-language retrieval
• Consolidation excludes pattern type from auto-splitting (only splits semantic and procedural)

0.2.0 - 2026-02-24

Added

• Auto-creation of memory types — session-end hook now creates procedural, semantic, and pattern memories automatically, not just episodic
  • Procedural: extracted from successful bash command sequences (build, test, lint, deploy workflows) when 2+ related commands run in a session
  • Semantic: technology stack detection from commands and file extensions, active module detection from modified file paths
  • Pattern: clustering of similar episodic memories (3+ with cosine similarity 0.40–0.95) with derived_from relations to source memories
• Automated consolidation — replaces the prompt-based "Consolidation Due" that Claude ignored
  • Merges near-duplicate memory pairs (cosine distance < 0.08, same type)
  • Detects patterns across all recent episodics
  • Deletes truly stale memories (importance < 0.1, access_count = 0, age > 60 days, excludes pattern/procedural)
  • Runs every 10 sessions, reports results in session header (e.g. "3 duplicates merged, 1 stale deleted")
• Enriched transcript parsing: bash command tracking with success/failure correlation via tool_use_id, files read, technology detection
• New database methods: getRecentEpisodicWithEmbeddings, getMemoriesByTypeWithEmbeddings, mergeMemories, findNearDuplicatePairs, deleteStaleMemories
• Shared utilities module (shared.ts): makeMemoryRecord, derivePatternTitle, STOP_WORDS
• Auto-created records tagged distinctively: auto-procedural, auto-semantic, auto-pattern

Changed

• Session-end now batches all embeddings in a single generateEmbeddings() call instead of sequential generateEmbedding() calls
• Session-end deduplicates new procedural/semantic records against existing memories before insertion (cosine distance < 0.05)
• CLAUDE.md instructions updated: documents auto-saved memory types, repositions manual memory_store as supplementary for insights automation might miss

0.1.3 - 2026-02-23

Improved

• Add behavioral reminder to session-start hook output — explicitly tells Claude that passive context is not a substitute for active memory_search
• Strengthen CLAUDE.md instructions: emphasize proactive recall during work (not just at task start), so users never have to say "check your memory"

0.1.2 - 2026-02-22

Fixed

• Handle nested message field in session transcript parsing

0.1.1 - 2026-02-21

Fixed

• Add error logging to embedding generation (previously silent failures)
• Fix unquoted shell variables in install script node -e blocks

Added

• Uninstall script (uninstall.sh)
• CHANGELOG

0.1.0 - 2025-02-21

Initial release of claude-code-memory.

Added

• MCP server with 9 memory tools (store, search, batch store, get, update, delete, relate, graph, health)
• SQLite database with WAL mode, FTS5 full-text search, and vec0 vector extension
• Hybrid search combining FTS5 text matching and vector similarity (384-dim local embeddings)
• 5 memory types: episodic, semantic, procedural, working, pattern
• Knowledge graph with 6 relation types and traversal up to depth 3
• Relevance scoring: text (0.5) + importance (0.2) + recency (0.2) + access frequency (0.1)
• Claude Code hooks: session-start (git signal context injection), session-end (episodic summaries), error-context (memory search on failures)
• Batch merge deduplication and contradiction auto-linking
• Importance decay for stale memories
• WAL checkpoint on database close
• Embedding model warmup for faster first-search performance
• Memory-init skill for bootstrapping project knowledge from codebase
• Memory-curator agent (Haiku-powered) for automated maintenance
• Install script with MCP server registration, skills, hooks, and permissions setup
• MIT license

v0.2.0

chore(changelog): update for release 0.2.0 with new features

v0.1.3

chore(release): bump version to 0.1.3

v0.1.2

chore(release): bump version to 0.1.2

v0.1.1 — Error logging, install fixes, uninstall script

Fixed

• Add error logging to embedding generation (previously silent failures)
• Fix unquoted shell variables in install script node -e blocks

Added

• Uninstall script (uninstall.sh)
• CHANGELOG

Full Changelog: v0.1.0...v0.1.1

v0.1.0 — Initial Release

claude-code-memory v0.1.0

First release of the Claude Code memory plugin — fully automatic, per-project cognitive memory.

Highlights

• MCP server with 9 memory tools (store, search, relate, graph, batch, health, etc.)
• Hybrid search — FTS5 full-text + vector similarity (local embeddings, 384-dim)
• SQLite storage with WAL mode, FTS5, and vec0 extension
• 5 memory types — episodic, semantic, procedural, working, pattern
• Knowledge graph — 6 relation types with traversal up to depth 3
• Claude Code hooks — automatic session-start (git signal context injection), session-end (episodic summaries), error-context (search on failures)
• Auto-maintenance — importance decay, deduplication with batch merge, contradiction auto-linking
• Embedding warmup for faster first-search performance
• Install script — one-command setup (./install.sh) with MCP registration, skills, hooks, and permissions
• Memory-init skill for bootstrapping project knowledge
• Memory-curator agent (Haiku-powered) for automated maintenance

Install

git clone https://github.com/d2a8k3u/claude-code-memory.git
cd claude-code-memory
./install.sh

Then restart Claude Code.