A typed knowledge graph with provenance, bitemporal search, contradiction detection, and 48 MCP tools. Available now as a Rust crate and MCP server. Works with Claude Desktop, Cursor, Hermes, and any MCP-compatible client.
Every other memory system is missing at least 5 of these 10 features. semantic-memory has all of them.
| Feature | semantic-memory | Zep / Graphiti | Letta / MemGPT | Mem0 | ChatGPT Memory |
|---|---|---|---|---|---|
| Typed graph edges (semantic / temporal / causal / entity) | Yes | Temporal only | No | No | No |
| Provenance on every fact (source, confidence, receipt) | Yes | No | No | No | No |
| Bitemporal search (what did I know on date X?) | Yes | No | No | No | No |
| Contradiction detection (decoder syndrome analysis) | Yes | No | No | No | No |
| Factor graph belief propagation over edges | Yes | No | No | No | No |
| RL-trained adaptive retrieval routing | Yes | No | No | No | No |
| Self-correcting (supersede facts when evidence changes) | Yes | No | Self-edit | No | No |
| Local-first (no cloud, no API keys, no telemetry) | Yes | Cloud | Optional | Cloud | Cloud |
| MCP server (48 tools, works with Claude Desktop, Cursor) | Yes | No | No | No | No |
| Auto-management (integrity checks, vacuum, re-embed, reconcile) | Yes | No | No | No | No |
Two crates work together: the library and the MCP server. The library is the knowledge graph engine. The MCP server exposes it to AI agents.
The core knowledge graph engine. Stores facts as typed nodes with heterogeneous edges (semantic, temporal, causal, entity). Every fact has provenance — source attribution, confidence score, and a receipt. Facts can be superseded when better evidence arrives. The system is bitemporal — it tracks when facts were valid and when they were superseded.
The MCP server wraps the library in 48 tools and 15 HTTP endpoints. It works with Claude Desktop, Cursor, Hermes, and any MCP-compatible client. Tool profile gating (lean / standard / full) keeps the tool list manageable — 34 daily-use tools in lean mode, 14 admin/audit tools hidden until needed.
Install the MCP server, add it to your client config, and you're done. Works with any MCP-compatible client.
Add this to your Claude Desktop config (Settings → Developer → Edit Config):
Add this to your Cursor MCP settings:
Three profiles control which tools are visible to the agent:
Switch with --tool-profile lean|standard|full. No recompilation needed.
AiDENs is the autonomous loop built on top of semantic-memory. It's in development — not published yet. This shows what the memory system enables when you add autonomous gap detection, LLM execution, and fact quality evaluation on top of it.
Real facts captured by the autonomous loop in the "autonomous" namespace. Each fact was generated by the model analyzing a detected gap, then quality-scored by the evaluation gate.
crates/aidens-autonomous/Cargo.toml file as a package manifest. Based on the file content and the project structure, this represents the "Brain" of the Loop — the autonomous loop executor that ties together gap detection, task generation, LLM execution, fact capture, and evaluation./Cargo.toml exists and is accessible, which contradicts the semantic memory integrity report. The discrepancy likely stems from a stale or incorrect reference rather than a missing file.semantic-memory and semantic-memory-forge crates, which AiDENs is tasked to wire and coordinate without duplicating.Cargo.toml of the llm-tool-runtime crate.The system is built from independent, reusable crates published on crates.io.