0.01ms writes. Sub-2ms search.
The first verified cross-model memory substrate for AI agents.
Writes
Recall@5 (1k nodes)
Multi-Agent Fidelity
500-Turn Retention
Same conversation. Same 7 days. Completely different outcomes.
A complete cognitive architecture, not just a vector database.
A unified memory layer accessible by any agent architecture, bridging the gap between Llama, Mistral, and GPT ecosystems.
Zero-loss context handover. Store with a reasoning model, retrieve with a coding assistant, or scale across thousands of specialized agents.
True cross-session persistence. Your agents maintain a permanent cognitive history that evolves as your user does.
Background LLM threads extract behavioral patterns and high-level insights from raw episodic logs.
Self-optimizing retrieval policies and schema evolution that adapts to user-specific interaction patterns.
C++ HNSW engines with Cross-Platform SIMD acceleration (AVX2/SSE/NEON). Millisecond search at 1,000,000 nodes.
Full system metrics: memory growth, recall latency, cluster health, schema evolution.
No cloud dependencies. All models run locally via Ollama. Data never leaves your machine.
Every interaction flows through 9 cognitive stages.
User sends a message
Detects user's cognitive intent
60s TTL deque captures it
Active context updated
Stored as timeline event
Embedded in 128D vector space
High-importance promoted
LLM generates insights
System self-optimizes
Every AI agent today suffers from goldfish memory. Each session starts from zero. No learning. No personalization. No continuity.
Recallix is a plug-and-play cognitive memory engine. A universal interop substrate with HNSW indexing, hardware-accelerated search, and self-evolving schemas for any AI.
Shared memory is the missing link for agent frameworks. Recallix becomes the standard substrate for collective AI intelligence.
Python AI brain for reasoning. C++ infrastructure for performance.