Memory that
understands context
Graph-aware semantic memory for AI agents. One database. Zero token waste.
$ openclaw-cortex capture --user "We use Memgraph for vectors and graph" --assistant "Good choice for unified storage"Captured [fact]: Memgraph chosen for unified vector + graph storage$ openclaw-cortex recall "database decision"[1] (92%) [fact] Memgraph chosen for unified vector + graph storage[2] (67%) [rule] Always validate schema on startup
Capabilities
Features
Everything you need to give your AI agent persistent, intelligent memory.
Semantic Recall
Vector similarity search across 768-dim embeddings with cosine distance.
Graph Traversal
Entity-seeded walks with Reciprocal Rank Fusion merge for richer context.
Smart Capture
Claude Haiku extracts structured memories from conversations automatically.
Temporal Versioning
Memories evolve over time with full version history and decay scoring.
Contradiction Detection
Conflicting facts are flagged with shared conflict groups and penalized.
Token-Aware Output
Recalled context is trimmed to fit your token budget — zero waste.
How It Works
Architecture
A clean layered design: your agent talks to the cortex, the cortex talks to the infrastructure.
All services are hot-swappable — swap Ollama for any OpenAI-compatible embedder, or Claude for any LLM via the gateway.
CLI
See It In Action
Capture memories from conversations and recall them with intelligent ranking — from the terminal or via the MCP plugin.
Capture
$ openclaw-cortex capture \
--user "We decided to use Memgraph for both \
vector search and graph traversal" \
--assistant "Solid choice — unified storage \
reduces operational overhead" \
--project "infra-decisions"
Embedding... done (768-dim)
Dedup check... no duplicates found
Captured [fact]: Memgraph chosen for unified vector + graph storage
confidence: 0.91 scope: project tags: [memgraph, storage, vectors]Recall
$ openclaw-cortex recall "database decision" \
--project "infra-decisions" \
--limit 3
Embedding query... done
Graph traversal... 4 entities seeded
[1] (92%) [fact] scope=project
Memgraph chosen for unified vector + graph storage
tags: [memgraph, storage, vectors]
[2] (67%) [rule] scope=permanent
Always validate schema on startup
tags: [startup, validation]
[3] (54%) [episode] scope=session
Evaluated Qdrant as alternative, ruled out
tags: [qdrant, storage, evaluation]Ready to give your agent memory?
Deploy in minutes with Docker. Works with any LLM via MCP or the CLI. Open source under MIT.