Features

Evidence-backed answers

Queries return answers with ranked memories, citations, confidence, and retrieval diagnostics so humans and agents can inspect why a fact was suggested.

Curated project memory

Memory Layer separates raw activity from durable memories. Curation and replacement proposals help keep current knowledge useful as the project changes.

Agent integrations

Use CLI commands, project-local instructions, watchers, and MCP tools to give coding agents project context without copying chat history by hand.

Multi-backend retrieval

Keyword, graph-aware, and optional embedding search can work together. Multiple embedding spaces can remain populated so switching active retrieval backends does not require recomputing everything.

Evaluations

The eval harness records artifacts and compares variants such as no-memory and full-memory. Treat metrics as evidence for bounded claims, not universal proof.

Next

Start with Quickstart, then read Concepts and Evaluations.