# Changelog
All notable changes to Ember MCP are documented here.
## [3.0.0] — 2026-02-26
### Complete Rewrite
Ember v3 is a ground-up rewrite focused on simplicity, speed, and reliability.
### Architecture
- **SQLite-only storage** — single `ember.db` file replaces FAISS index + JSON files + stats.db. No more `faiss-cpu` dependency.
- **4-tier memory** — working → session → relational → glacier with TTL-based expiry and access-based promotion.
- **FastMCP transport** — compatible with MCP SDK v1.26.0+.
- **FTS5 full-text search** — automatic fallback when semantic search returns empty results.
### New Features
- **CORAL checkpointing** — save and resume complex task state mid-session. Detect cognitive overload from repeated context switches.
- **Consolidation engine** — auto-promotes frequently accessed memories between tiers based on access patterns.
- **Partial ID support** — all mutation tools (update, delete, discard, set_status) accept 8-character short IDs.
- **Dual search** — semantic similarity with FTS5 fallback on every search tool (recall, auto, deep_recall, graph_search).
### Breaking Changes
- **Storage location** moved from `~/.ember/` to `~/.ember-v3/`. Run `ember-mcp migrate` to migrate existing data.
- **Python 3.11+** required (was 3.10+).
- **`faiss-cpu` removed** from dependencies. FAISS is no longer used.
- **`aiosqlite`, `aiofiles`, `filelock`, `pydantic` removed** from dependencies.
- **`ember-mcp run`** no longer needed — just `ember-mcp` starts the server.
### Migration
Run `ember-mcp migrate` to migrate v1/v2 memories, edges, and re-generate embeddings for the new format.
---
## [0.2.0] — 2026-02-21
### New Tools
- **`ember_compact`** — AI-powered memory compaction. Analyze mode identifies stale/shadowed candidates; execute mode replaces content with an LLM-generated summary and re-embeds the vector.
- **`ember_actionable`** — List embers with active task status (`open` or `in_progress`), sorted by priority.
- **`ember_set_status`** — Update the task status of any ember (`open`, `in_progress`, `done`, or clear).
### Enhancements
- **`ember_store`** — New `status` parameter for task tracking (`open`, `in_progress`, `done`). New `edges` parameter for typed knowledge graph edges (`depends_on`, `child_of`, `context_for`). Invalid `importance` or `status` values now return an error string instead of silently coercing to defaults.
- **`ember_graph_search`** — New `edge_types` parameter to filter BFS traversal by edge type.
- **`ember_auto`** — Session continuity boost: session-sourced embers receive a 1.5x HESTIA score multiplier; the most recent session ember receives 2.0x. Candidate pool expanded from 5 to 10 before boosting.
### Models
- `Ember` model: new fields `status`, `is_compacted`, `original_content_length` (all optional with defaults — no migration required for existing embers).
- New constants: `VALID_EDGE_TYPES`, `USER_EDGE_TYPES`.
### Security
- **Path traversal fix (C1):** `ember_deep_recall` now restricts `source_path` reads to within `Path.home()`. Paths outside this boundary are skipped with a `[Skipped: path outside home directory]` annotation.
### Registry
- Added `server.json` for MCP Registry listing at `registry.modelcontextprotocol.io`.
## [0.1.0] — Initial release
- 14 MCP tools: `ember_store`, `ember_recall`, `ember_read`, `ember_deep_recall`, `ember_learn`, `ember_contradict`, `ember_list`, `ember_delete`, `ember_inspect`, `ember_auto`, `ember_save_session`, `ember_drift_check`, `ember_health`, `ember_recompute_shadows`, `ember_explain`.
- FAISS-based semantic search with 16-cell Voronoi partitioning.
- Shadow-Decay framework with HESTIA scoring.
- SQLite knowledge graph with `related`, `shadow`, and `supersedes` edge types.
- Local-only storage at `~/.ember/` — no cloud dependencies.
