How sensesvoid/sqac works
SQAC turns taught facts, imported datasets, and project knowledge into portable .sqac cartridges, then retrieves a small set of relevant facts or skills for an LLM at query time. Its broader product also supports conversation offloading and session continuity, cartridge routing, code-graph exploration, and CLI, MCP, and HTTP access. The README documents several capabilities whose implementation excerpts were not sampled; those links are shown only where the documented workflow supports them.
Main entry points: store.py Fact store · sqac_mcp.py MCP tools · server.py HTTP API · autobuild.py Project auto-build
Components 21
Knowledge authoring
- Project auto-build
sqac/autobuild.py - Realtime tracking
sqac/autobuild_watcher.py - Dataset ingestion
sqac/ingest.py - Skill routing
sqac/skills.py - Code parser
sqac/graph/parser.py
Cartridge runtime
- Fact store
sqac/store.py - Address encoding
sqac/encoder.py - Semantic encoder
sqac/static_encoder.py - SQAC cartridge
- Cartridge rack
sqac/rack.py - Thought injection
sqac/injection.py
Context lifecycle
- Context offloader
sqac/offloader.py - Session continuity
sqac/continuity.py - Checkpoints
sqac/checkpoint.py
Access and exploration
- CLI
sqac/cli.py - MCP tools
sqac_mcp.py - HTTP API
sqac/server.py - Code graph
sqac/graph/graph.py - Graph MCP
sqac/graph_mcp.py
Other components
- User / application
- LLM
Connections 28
- User / application → CLI: teaches / queries
- CLI → Fact store: writes / searches
- User / application → Project auto-build: selects project
- Project auto-build → Fact store: extracts knowledge
- Realtime tracking → Project auto-build: tracks changes
- User / application → Dataset ingestion: provides dataset
- Dataset ingestion → Fact store: adds records
- Fact store → Address encoding: encodes addresses
- Fact store → Semantic encoder: semantic matching
- Fact store → SQAC cartridge: saves / loads
- Cartridge rack → SQAC cartridge: routes cartridges
- Fact store → Skill routing: routes skill hits
- Skill routing → Thought injection: selects procedure
- Fact store → Thought injection: provides retrieved context
- Thought injection → LLM: injects procedure
- Fact store → LLM: returns relevant facts
- User / application → MCP tools: requests memory tools
- MCP tools → Fact store: reads / writes
- User / application → HTTP API: sends API requests
- HTTP API → Fact store: searches / teaches
- HTTP API → Cartridge rack: routes named cartridge
- User / application → Context offloader: provides conversation
- Context offloader → Fact store: offloads / recalls turns
- Session continuity → Checkpoints: records session state
- MCP tools → Session continuity: manages sessions
- User / application → Code parser: selects codebase
- Code parser → Code graph: extracts symbols and edges
- Graph MCP → Code graph: queries code graph