Because the store lives under the Git project rather than inside one tool's session, what yesterday's Claude Code session worked out is available to today's Cursor session — and the review gate keeps that store from filling up with everything anyone ever typed.
A local-first shared memory layer for AI coding agents. SQLite is the canonical store and Orama handles search; project identity comes from the real Git root, so the memory belongs to the repository rather than to whichever tool you happened to use.
- `memorix_project_context` — the default brief an agent starts from: current project facts, start-here files, selected evidence, cautions about stale or suspect memories, and lens-specific verification hints, with the task lens inferred from the request
- Retrieval that expands only when needed: `memorix_search`, `memorix_detail` for a named missing fact, `memorix_context_pack` for a task-specific packet, and `memorix_graph_context` for an evidence-governed graph overview
- Writing memory: `memorix_store`, `memorix_store_reasoning` for design rationale with alternatives and trade-offs, `memorix_search_reasoning`, `memorix_timeline` and `memorix_resolve`
- Sessions: `memorix_session_start` — which is also how an HTTP-mode agent binds the active repository — plus `memorix_session_end` and `memorix_session_context`
- `memorix_evidence` and `memorix_feedback` for the evidence and feedback lifecycles, `memorix_codegraph_status` for index counts, `memorix_media` as the compact media companion available in every profile, and `memorix_promote` for turning durable knowledge into reusable skills
- Advanced tools registered only in the wider profiles: `memorix_knowledge` for the review-gated knowledge workspace, and `memorix_compaction_checkpoint` for explicit continuity inspection
- Git Memory turning commits into searchable engineering facts, orchestration for multi-agent work with handoffs, file locks and review loops, a local dashboard, and `memcode`, a bundled terminal agent reading the same memory
Node.js 22.18.0 or newer, and Git — project identity is derived from the real Git root, so a directory outside a repository has no identity to attach memory to. Install with `npm install -g memorix`, then `memorix setup --agent claude --global`, substituting the agent you use; setup covers Claude Code, Codex, Copilot, Cursor, Gemini CLI, OpenCode, Windsurf, Kiro and others, choosing plugin, MCP, rules, hooks or skills depending on what each supports. `memorix init` is optional and writes TOML configuration to `~/.memorix/config.toml` globally or `memorix.toml` per project. Model keys are optional: without them, memory formation and embeddings are skipped and local full-text retrieval still works. Long-term memory is deliberately not automatic — a record has to be created and then approved, and only a manually created user item marked portable can be considered in another local project.
One command — npm install -g memorix