Memorix is a local-first project memory that several AI coding agents share. Memory is scoped to the Git project rather than to one chat or one tool, so facts, fixes, and decisions recorded in a Claude Code session are available to Codex, Cursor, or another agent in the next session.
Agents reach the memory through the interface each one supports: an MCP server, plugin or extension packages, project rules, hooks, skills, a CLI, a TypeScript SDK, or memcode, a terminal coding agent bundled with Memorix. A setup command writes the right integration for each agent, and a doctor command checks and repairs it.
SQLite is the canonical store. Small projects search through an in-process Orama index, larger ones through a SQLite FTS5 index plus an optional local LanceDB semantic index, and both indexes are rebuildable. LLM-backed memory formation and embeddings are optional.
| Layer | Stores |
|---|---|
| Observation Memory | facts, gotchas, fixes, implementation notes |
| Reasoning Memory | rationale, alternatives, constraints, risks |
| Git Memory | commit-derived engineering facts |
| Code Memory | files, symbols, import edges, and memory-to-code freshness |
| Curated Long-term Memory | reviewed episodic events, stable facts, or reusable procedures with evidence |
| Compact Continuity | recent compaction summaries or lifecycle markers |
Search is project-scoped by default and boosts Git Memory for "what changed" questions and Reasoning Memory for "why" questions. Beyond storage, Memorix builds task-shaped context briefs, coordinates multi-agent work with handoffs, file locks, and review loops, and serves a local dashboard for browsing and cleaning up memory.
Features
- Memory Autopilot: builds a compact task brief with start files, current memory, cautions, and verification steps, weighted by task type
- Git Memory: turns commits into searchable engineering facts about what changed, where, and why
- Reasoning Memory: keeps design rationale, alternatives, trade-offs, and risks beyond a single chat
- Code Memory: versioned code snapshots with TypeScript and JavaScript symbols, relations, and freshness checks against stored memories
- Curated long-term memory: reviewed episodic, semantic, and procedural records with source evidence, kept separate from automatic capture
- Knowledge workspace: review-gated claims, Markdown pages, and project workflows that proposals never overwrite silently
- Agent setup and repair: one command installs MCP, rules, hooks, skills, or plugins per agent, and a doctor command repairs stale entries
- MCP tool profiles: micro, lite, team, and full profiles expose 9 to 48 tools to keep schemas small
- Orchestration: task planning, worker handoffs, file locks, messages, verification gates, and review loops across parallel agents
- memcode: a bundled terminal coding agent that reads and writes the same project memory
- Dashboard and HTTP service: a local web UI and shared MCP endpoint for browsing, deduplication, consolidation, and retention, deployable with Docker
- Multi-device sync: opt-in, privacy-filtered relay of observation events through GitHub JSONL, a filesystem, S3, or Postgres
