Memorix

Memorix is a local-first memory layer that gives AI coding agents one shared, searchable project memory instead of a separate context per chat window or tool. The memory lives under the Git project, so a session with one agent can pick up what another worked out yesterday, and switching editor or terminal does not reset what is known.

SQLite is the canonical store and Orama handles search; language-model-backed memory formation and embeddings are optional, and without model keys the system still works on local full-text retrieval. Search is project-scoped by default, boosts commit-derived records for "what changed" questions and rationale records for "why" questions, and can be widened across projects. Three entry points sit on the same store: a command-line interface that binds to the current Git project without an MCP session, a stdio MCP bridge for editors and agents, and an HTTP service for a shared endpoint, dashboard, or container deployment.

Memory is layered rather than a single pile of notes, and durable records are review-gated: a candidate is created, qualified, and then explicitly approved before it can anchor a task brief. Only a manually created or user-confirmed item marked portable may cross into another local project.

Layer	Stores	Answers
Observation Memory	facts, gotchas, fixes, implementation notes	How does this work?
Reasoning Memory	rationale, alternatives, constraints, risks	Why did we choose this?
Git Memory	commit-derived engineering facts	What changed and where?
Code Memory	files, symbols, import edges, memory-to-code freshness	Which code should I inspect first?
Curated Long-term Memory	reviewed episodic events, stable facts, reusable procedures with evidence	What should this agent still know later?
Compact Continuity	host-native compact summaries and lifecycle markers	What survived the last context compaction?

A setup command wires each agent through whatever surface it already supports — plugin or bundle package, extension, MCP config, project rules, hooks, or skills — with stdio MCP as the default transport. The MCP surface is sized by profile: a compact core of 9 tools, 20 for the normal memory and session surface, 28 with coordination tools, and 47 including advanced and compatibility tools.

Memorix also coordinates multi-agent work through task planning, worker handoffs, file locks, messages, verification gates, and review loops; multi-worker runs create task worktrees and merge successful branches back. It bundles memcode, a terminal coding agent reading the same project memory as the MCP-connected agents, and ships a TypeScript SDK.

Features

• Memory Autopilot: builds a bounded, task-lensed workset of start files, current memory, source-backed knowledge, cautions, and verification, with JSON receipts of what was selected or omitted
• Git Memory: turns commits into searchable engineering facts through a command or a git hook
• Reasoning Memory: keeps design rationale, alternatives, trade-offs, and risks beyond the chat they were decided in
• Code state and code graph: versioned local code snapshots, source-backed TypeScript and JavaScript symbols and relations, file and symbol links, and freshness checks
• Knowledge workspace: review-gated source-backed claims, Markdown pages, and canonical workflows where proposals never silently overwrite reviewed pages
• Agent setup and repair: one setup path per agent plus doctor and repair commands for stale or missing configuration
• Orchestration: task planning, handoffs, file locks, messages, verification gates, and optional worktree isolation per worker
• Controlled media: opt-in local import outside the Git worktree, hashed and attachable as memory, with vision analysis capped at 20 MiB
• memcode: a bundled terminal coding agent that edits, runs commands, resumes sessions, and switches models against the same memory pool