curl -fsSL https://raw.githubusercontent.com/quaid-app/quaid/main/scripts/install.sh | shThen build your first memory and connect it to Claude Code:
quaid init ~/memory.dbquaid collection add notes ~/notes --db ~/memory.dbQUAID_DB=~/memory.db quaid serveThree commands. No API keys. No cloud. Full install options →
For developers
One binary. One file. SQLite-backed full-text search and a local vector index, with a CLI you can drive from any shell. Ten minutes from install to a memory that answers your first question.
For agents
Twenty-four memory_* tools, knowledge-graph path output, daemon lifecycle commands (quaid daemon), quaid status, and an opt-in HTTP/SSE MCP transport — all available in the current published release (v0.22.2). Same surface humans use; same data, same contracts. Privacy-safe gap logging, OCC-protected writes, and a sensitivity contract that keeps raw queries from ever leaving the machine by accident.
Memories accumulate. Nothing is rewritten. A head pointer resolves queries to current truth, but the full history is always accessible. You can ask what you believed before, trace how a view evolved, and audit the chain at any time.
Markdown defines what’s true. SQLite stores it deterministically. Embeddings only index it. The index is rebuildable; the truth is not. Every layer is inspectable, and you can read every page with a text editor.
Compute and storage on your machine. The embedding model ships in the binary on the airgapped channel. No container, no Python runtime, no service mesh, no API keys. Drop one statically-linked file on a laptop, a Pi, or an air-gapped workstation and it works the same.
Read, write, graph, intelligence, collections, system.
Available in the current published release (`v0.22.2`).
Ingest, query, maintain, enrich, briefing, alerts, research, upgrade.
No Docker, no Python, no cloud. Background daemon (`quaid daemon install`) available in `v0.22.2`.
The model runs on your CPU. Nothing transits your network.
The benchmark gap is real, and it's architectural
On LongMemEval, Quaid scores 0.0% today. Mem0 v3 scores 93.4%. On LoCoMo, Quaid scores 0.1%; Mem0 v3 scores 91.6%. The gap isn’t a retrieval failure - Quaid stores conversation turns whole, and the benchmark expects extracted facts. Phase 5 ships the local SLM that closes it: Phi-3.5 Mini, fully airgapped, downloaded once. Target after implementation: LongMemEval > 40%, LoCoMo > 40%. See the roadmap →
Workflows ship inside the binary, not in a prompt
Most agent tooling ships workflow logic in the cloud or stuffed into a prompt. Quaid ships eight skills as Markdown files inside the static binary - ingest, query, maintain, enrich, briefing, alerts, research, upgrade. They extract to ~/.quaid/skills/ on first run, and any SKILL.md dropped in your working directory overrides the default. The agent reads what’s on disk; you read what the agent reads.
Memory that follows your editor
Attach any Obsidian vault as a collection. Quaid watches for edits with a 1.5 s debounce and reconcile-backed flushes, so the memory an agent reads is always the same memory you just updated. Hybrid search across FTS5 and a local BGE vector index runs on every write. (Unix/macOS/Linux only - v0.9.6+)
Three commands. No API keys. No cloud.
# Initialize a memoryquaid init ~/memory.db
# Attach your notes as a live-sync collectionquaid collection add notes ~/notes --db ~/memory.db
# Expose Quaid's 24 MCP tools over stdio (v0.22.2; opt-in HTTP/SSE transport via --http flag)QUAID_DB=~/memory.db quaid serveNext: Build your first memory · Connect Claude Code · Why Quaid