omega-memory/omega-memory releases

Features

• Reranker precision selection -- Set OMEGA_RERANKER_PRECISION=int8 to use the quantized model (~571 MB on disk, ~650 MB RSS) instead of the FP32 default (~2.3 GB on disk, ~2.5 GB RSS). Also available via download_model(precision="int8"). Auto-detect finds whichever variant is installed.

Upgrade

pip install --upgrade omega-memory

Fixes

• fix: download_model() 404 on bge-reranker-v2-m3 — BAAI upstream repo no longer ships ONNX weights. Repointed to onnx-community/bge-reranker-v2-m3-ONNX (HuggingFace's canonical ONNX conversion). Also added model.onnx_data sidecar and fixed download verification logic. Closes #53.
• security: depersonalize codebase and harden credential handling

Upgrade

pip install --upgrade omega-memory

Fix: Pro activation works from bare omega-memory install

omega activate now works without requiring the Pro wheel to be pre-installed:

• Validates the license key directly against the activation API
• Auto-downloads and installs the Pro package
• Single command from install to fully working Pro

Previously, omega activate failed with "OMEGA Pro is not installed" because it required a package that only exists after Pro setup, creating a chicken-and-egg problem.

Upgrade

pip install --upgrade omega-memory
omega activate OMEGA-PRO-<your-key>

Free-tier protocol fallback with pro-reminder

What's Changed

Added

• Pro-reminder system with graduated urgency based on memory count
• GitHub star ask in CLI status card, setup, and welcome briefing
• Tool gate: upgrade CTA when free users call any of 85 Pro-only tools (was bare "Unknown tool" error)

Changed

• Periodic reminder interval reduced (less frequent, more contextual)
• Milestone warnings now range-based (1,000+/1,800+) instead of exact counts
• Session stop hook shows memory count vs limit at graduated frequency
• All LLM-facing upgrade notices use "Tell the user:" instruction format so agents relay to humans
• Degradation warnings start earlier with approaching-limit context

Fixed

• omega_platform.license import paths corrected for Pro detection

Add MCP registry verification tag for official listing.

Remove telemetry phone-home. Local-only counters retained for CLI display. No data sent to any server.

Quality degradation after 2,000 memories (keyword-only search for free tier). CLI email capture at milestones.

Upgrade prompts, telemetry, omega upgrade command. See README for details.

Republish of v1.3.0 changelog (v1.3.0 was already claimed on PyPI from an earlier partial release).

Highlights

• Condensed mode (~80% context token savings, on by default)
• Behavioral learning engine (tool preferences, git style, workflow sequences)
• Advisory engine (context-aware suggestions for edits, errors, deploys)
• CrewAI integration, Obsidian export, stats card

Fixed

• _get_store() crash (#48): handlers.py imported non-existent omega.store module, breaking omega_query(mode="browse"), omega_stats, omega_reflect, and memory link/flag/supersede actions
• Recursive omega_call to meta-tools
• 5 test failures (#44)

Changed

• omega doctor --client now supports 8 clients (was only claude-code)
• export/import CLI subcommands removed (use omega export-obsidian)

See CHANGELOG.md for full details.

Highlights

• Condensed mode (~80% context token savings, on by default)
• Behavioral learning engine (tool preferences, git style, workflow sequences)
• Advisory engine (context-aware suggestions for edits, errors, deploys)
• CrewAI integration, Obsidian export, stats card

Fixed

• _get_store() crash (#48): handlers.py imported non-existent omega.store module, breaking omega_query(mode="browse"), omega_stats, omega_reflect, and memory link/flag/supersede actions
• Recursive omega_call to meta-tools
• 5 test failures (#44)

Changed

• omega doctor --client now supports 8 clients (was only claude-code)
• export/import CLI subcommands removed (use omega export-obsidian)

See CHANGELOG.md for full details.

What's New

• Batch store: omega_store now accepts an items array for batch memory storage
• Cache optimization: STABLE_EVENT_TYPES for prompt cache-friendly content ordering
• Unified query mode: Cross-search memories and knowledge documents with mode="unified"
• Constraint management: omega_maintain now supports list_constraints, check_constraint, save_constraints
• Insight synthesis: omega_maintain action synthesize_insights for generating system insights
• Embedding backfill: omega_maintain action backfill_embeddings to fill missing vectors
• Graph stats: omega_stats now supports graph_stats action
• Extended event types: Oracle, entity, and behavioral pattern event type constants (graceful degradation for core users)

Full Changelog: https://github.com/omega-memory/omega-memory/compare/v1.1.0...v1.2.0

OMEGA v1.1.0

OMEGA is stable and production-ready. This release adopts 1.x semver to signal API stability.

Changed

• Version bump to 1.x: Semantic versioning 1.x signals stable API
• Synced core files from upstream (preferences.py, test_types.py, plugins.py, tool_schemas.py, types.py)
• Added trace query mode for session tool call timeline

Fixed

• Security: Removed pro-only files (protocol.py, license.py) that were accidentally included in the repository

Documentation

• Updated SECURITY.md supported versions
• Updated CHANGELOG.md

Full Changelog: https://github.com/omega-memory/omega-memory/compare/v0.12.0...v1.1.0

What's Changed

Added

• Hooks in wheel: Claude Code hooks now ship inside the pip package; previously pip-installed users got broken hook errors because fast_hook.py was excluded
• omega hooks CLI: New subcommand with three modes:
  • omega hooks setup — auto-configures ~/.claude/settings.json with correct paths
  • omega hooks path — prints the hooks directory (machine-readable)
  • omega hooks doctor — diagnoses broken paths in hook config

Fixed

• Hooks excluded from wheel due to missing __init__.py in src/omega/hooks/

Full Changelog: https://github.com/omega-memory/omega-memory/compare/v0.11.1...v0.12.0

Fixed

• Fixed browse and flagged-memory queries returning integer row IDs instead of string node IDs, causing TypeError crashes when formatting results.
• Fixed 7 CLI commands crashing with ImportError on community edition when pro-only modules (omega.knowledge, omega.cloud, omega.license, omega.server.http_server) are not installed.

Added

• 91 new tests covering bridge helper functions (_extract_facts, _auto_relate, _detect_and_supersede, _split_atomic_facts) and handler actions (omega_reflect, omega_memory, omega_stats, omega_browse).
• Updated integration test assertions to match current tool and handler counts.

Full Changelog: https://github.com/omega-memory/omega-memory/compare/v0.11.0...v0.11.1

What's New

New Tools

• omega_reflect — Self-evaluation tool for reviewing recent memories
• omega_consult_gpt / omega_consult_claude — Multi-LLM consultation

Memory Intelligence

• Typed relationship edges — related, supersedes, evolution, same_entity, temporal_cluster
• Contradiction supersession — Conflicting memories auto-superseded
• Atomic fact splitting — Compound memories decomposed for better recall
• Fact extraction — Auto-extracted retrieval keys from high-value memories

Stability

• SQLite thread executor — Prevents GIL+GC SIGSEGV under concurrent access
• RSS watchdog — Graceful exit before OOM crash
• Rotating log files — 5 MB cap, 3 rotations

Other

• Extended TTLs — Decisions/insights now permanent, session summaries 90 days
• 14 tools (was 12) — Enhanced schemas for omega_query, omega_memory, omega_stats
• embedding.py rename — graphs.py → embedding.py with backward-compat shim

Full changelog: https://github.com/omega-memory/omega-memory/blob/main/CHANGELOG.md

Fixes

• Auto-repair broken hook paths — omega setup now detects and replaces hooks with incorrect paths instead of adding duplicates. Fixes the issue where setup pointed hooks to Python framework directories that don't contain hook files.

Fixes

• Fix hook path resolution — omega setup was generating hook commands pointing to the wrong path on pip installs (e.g. /Library/Frameworks/Python.framework/.../hooks/fast_hook.py which doesn't exist). Hook files are now bundled inside the package.
• Add omega activate and omega license commands (from v0.10.10)

What's New

• omega activate — Activate a Pro license key from the CLI
• omega license — View Pro license status or deactivate

These commands enable Pro customers to activate their license directly via the CLI:

omega activate OMEGA-PRO-your-key-here

Fixed

• Test fixture: CoordinationManager in conftest.py now passes cloud_sync=False to prevent cloud sync attempts during tests.
• __init__.py version string now matches pyproject.toml (was stuck at 0.10.6).

Also includes (0.10.7, previously untagged)

• MCP server now sends instructions during initialization for automatic omega_welcome() usage.
• Synced core files and added pro-only test guards.

What's new

• Automatic MCP instructions: The MCP server now sends instructions during initialization that tell Claude to use memory tools proactively. Claude Desktop users no longer need to manually configure project instructions.

What's Changed

Added

• WAL checkpoint management: TRUNCATE on init, PASSIVE on close, periodic every N writes (configurable via OMEGA_WAL_CHECKPOINT_INTERVAL). Prevents unbounded WAL growth under multi-process contention.
• Auto-backup on startup when most recent backup is >24h old (max 5 rotated in ~/.omega/backups/)
• _run_sql() retry wrapper for individual SQL statements under lock contention
• Thread-safe query cache with _cache_lock protecting all reads/writes
• Smart partial cache invalidation (trigram overlap >= 0.20 vs full wipe)

Fixed

• find_similar() now filters expired and superseded memories (2x over-fetch to maintain result count)

Tests

• 733 tests passing

Full Changelog: https://github.com/omega-memory/omega-memory/compare/v0.10.5...v0.10.6

What's Changed

Added

• Windows/WSL installation guide in README with step-by-step setup and 5 WSL-specific gotchas (closes #19)
• db_utils.py shared SQLite retry helpers for database-locked error recovery

Fixed

• SQLite CLI connections now use timeout=30 to prevent WAL lock hangs; doctor checks use timeout=5
• SESSION_START surfacing thresholds broadened for better context at startup
• Three noise filters in auto-capture: infrastructure events, zero-token outcomes, JSON blob decisions
• Cross-session dedup exception for decisions
• Shortened capture output messages for cleaner hook output
• Lowered dedup thresholds (session summaries 0.95→0.75, decisions 0.85→0.80)
• exc_info=True on all bridge.py logger.error() calls

Tests

• 733 tests passing

Full Changelog: https://github.com/omega-memory/omega-memory/compare/v0.10.4...v0.10.5

What's Changed

Features

• --json flag fully tested on query, status, timeline, stats, and activity commands (closes #9)

Improvements

• Session summary TTL changed from SHORT_TERM to EPHEMERAL (1h) to prevent accumulation
• Generic memory type TTL changed from LONG_TERM to SHORT_TERM (1d) for faster expiry
• consolidate prune_days default lowered from 30 to 14 days
• Added exc_info=True to all 30 logger.error() calls in handlers for better tracebacks
• Eliminated all datetime.utcnow() deprecation warnings (Python 3.12+)

Tests

• 733 tests passing (up from 729)
• 4 new --json output tests for CLI commands
• Zero deprecation warnings from test suite

Full Changelog: https://github.com/omega-memory/omega-memory/compare/v0.10.3...v0.10.4

Fixes

• Crash prevention: Safe store initialization with proper double-checked locking
• Memory quality: Raised cosine similarity threshold (0.50 → 0.60) to reduce irrelevant results
• Memory quality: Raised minimum content length (40 → 80) to reduce low-value captures
• Memory quality: Hash-fallback embedding warning surfaced in query metadata
• Hardening: Added missing DB indexes (last_accessed, event_type+access_count, ttl_seconds)
• Hardening: Atomic import with BEGIN EXCLUSIVE transaction to prevent empty-DB windows

Patch release: fix welcome enrichment, doctor test, CI, social preview. Enables MCP Registry publishing.

What's Changed

MCP is now an optional dependency. This makes omega-memory lighter for users who only need the Python API, while keeping full MCP support available via an install extra.

Install

# For MCP clients (Claude Code, Cursor, Windsurf, Zed)
pip install omega-memory[server]

# For Python scripts, CI/CD, automation (no MCP server)
pip install omega-memory

Highlights

• pip install omega-memory no longer pulls in the MCP package (~50MB lighter, no server process)
• pip install omega-memory[server] works identically to previous versions
• omega setup --hooks-only configures hooks without MCP server (~600MB RAM saved per session)
• 9 new functions in the public Python API (batch_store, create_reminder, deduplicate, etc.)
• Clear error message if you try to run the MCP server without the [server] extra

Migration

If you use OMEGA with Claude Code, Cursor, or any MCP client, update your install:

pip install omega-memory[server]

See the full CHANGELOG for details.

Tool Consolidation: 26 -> 12 MCP Tools

Reduces context window overhead by ~50% while preserving all capabilities via action-discriminated composite tools.

Changes

• 7 standalone tools: omega_store, omega_query (+timeline mode), omega_welcome, omega_protocol, omega_lessons, omega_checkpoint, omega_resume_task
• 5 composite tools: omega_memory (edit/delete/feedback/similar/traverse), omega_profile (read/update/list_preferences), omega_remind (set/list/dismiss), omega_maintain (health/consolidate/compact/backup/restore/clear_session), omega_stats (types/sessions/digest)
• All 26 original tool names remain as backward-compatible handler aliases
• 671 tests passing

Why This Matters

• LLM accuracy drops significantly with tool count (TaskBench: 96% with 1 tool, 25% with 8 tools)
• VS Code enforces a 128-tool hard limit per LLM request
• Fewer tools = faster tool selection, lower context overhead, better agent performance

Fix demo GIF rendering on PyPI — use markdown image syntax instead of HTML img tag (PyPI sanitizer strips HTML images).

Fix demo GIF rendering on PyPI (use absolute URL for image).

What's New

• Compact MCP responses — reduced token usage across all tool outputs
• Contradiction detection — new decisions automatically supersede conflicting old ones
• Atomic fact splitting — complex memories broken into retrievable atomic facts
• Corpus hygiene — periodic dedup of near-duplicate memories
• Demo GIF — animated terminal demo in README
• README polish — comparison table, badges, star history

Full changelog: https://github.com/omega-memory/core/blob/main/CHANGELOG.md

• Fix MCP Registry ownership validation (visible tag in README)
• Homepage links to omegamax.co on PyPI

• Homepage now links to omegamax.co on PyPI
• Documentation URL added (omegamax.co/how-it-works)
• MCP Registry ownership tag added to README

Added

• omega status --json — machine-readable JSON output for scripted access to memory count, DB size, model status, and vector search availability
• omega export — export memories to a JSON file, with optional --type filter (e.g. omega export --type decision decisions.json)
• omega import — import memories from a JSON file, with optional --clear to replace existing data

Changed

• CONTRIBUTING.md — expanded with full dev setup, test commands, code style guide, MCP server testing instructions, project structure, and PR process

Closes #2, #4, #5

Full Changelog: https://github.com/omega-memory/core/compare/v0.7.3...v0.8.0

Fixed

• Welcome briefing false positive — model-missing warning checked backend activation state (lazy-loaded, always None at welcome time) instead of model files on disk.

Full Changelog: https://github.com/omega-memory/core/compare/v0.7.2...v0.7.3

Fixed

• Missing model warning — when the ONNX embedding model is not downloaded, OMEGA now shows a clear actionable error ("Run 'omega setup' to download the model") instead of silently falling back to degraded text-only search. Warnings appear in CLI queries, model loading, and the MCP welcome briefing.

Closes #3

Full Changelog: https://github.com/omega-memory/core/compare/v0.7.1...v0.7.2

Fixed

• MiniLM model download — tokenizer and config files were fetched from the wrong HuggingFace path (onnx/ subdirectory instead of repo root), causing 404 errors during omega setup for first-time users without the bge model.

Full Changelog: https://github.com/omega-memory/core/compare/v0.7.0...v0.7.1

Multi-Client Setup

omega setup --client cursor|windsurf|zed writes MCP config to each editor's native config file. Claude Code remains the default with full hooks and instruction injection.

Added

• Multi-client setup — Cursor (~/.cursor/mcp.json), Windsurf (~/.codeium/windsurf/mcp_config.json), Zed (~/.config/zed/settings.json)

Changed

• Setup auto-detect now lists all supported clients when Claude Code is not found

Full Changelog: https://github.com/omega-memory/core/blob/main/CHANGELOG.md

Fixed

• SQLite lock contention — increased busy_timeout from 5s to 30s and added retry-with-backoff on all write paths. Fixes "database is locked" errors when multiple Claude Code sessions share the same omega.db.

Added

• Entity auto-capture — resolve_project_entity() wired into bridge.auto_capture() for automatic entity scoping.
• Smithery.yaml — configuration file for Smithery.ai directory listing.
• Demo GIF — animated terminal demo in README showing cross-session memory recall.

Changed

• README restructured: leads with problem statement, demo GIF, and examples section.

Upgrade: pip install --upgrade omega-memory

Forgetting Intelligence

• Forgetting Audit Trail — every deletion logged with reason (TTL, LRU, consolidation, feedback, user). New omega_forgetting_log tool to query the log.
• Decay Curves — old unaccessed memories rank lower in search results. Preferences and errors exempt. Floor at 0.35.
• Conflict Detection — contradictions auto-detected on store. Decisions auto-resolve (newest wins), lessons get flagged for manual review.

26 MCP tools total (+1). 31 new tests for forgetting intelligence features.

Full Changelog: https://github.com/omega-memory/core/blob/main/CHANGELOG.md

pip install --upgrade omega-memory

Initial Open Source Release

OMEGA — persistent memory for AI coding agents. First public release under Apache-2.0.

Core Memory System

• 25 MCP tools for storing, querying, and managing long-term memory
• Semantic search via bge-small-en-v1.5 embeddings + sqlite-vec
• FTS5 full-text search for exact phrase queries
• Graph relationships (related, supersedes, contradicts) with BFS traversal
• Memory compaction and consolidation for long-term hygiene
• Timeline views and session-scoped queries
• Context virtualization (checkpoint/resume)

Auto-Capture & Surfacing

• Hook system for automatic context capture and memory surfacing
• Session start/stop lifecycle with welcome briefing
• Pre-edit memory surfacing with file-extension-aware re-ranking
• Auto-capture of decisions, lessons, and error patterns

Storage & Security

• SQLite + sqlite-vec backend (local-first, no cloud required)
• Optional AES-256-GCM encryption at rest (pip install omega-memory[encrypt])
• Database created with 0o600 permissions
• Export/import for backup and restore

Developer Experience

• omega setup — one-command installation
• omega doctor — health diagnostics
• omega query/store/remember — CLI access to memory
• Plugin architecture via entry points for extensibility

Install: pip install omega-memory