Distributed neural simulation with MPI for Brian2 models
Agent memory that learns from experience
Stage
Startup
Startup
Founded
2026
2024
Based in
GB
US
Pricing model
Open source
Freemium
Pricing
Not given
Self-hosted: Free, MIT licensed. Hindsight Cloud: Pay-as-you-go with free credits to start. Retain (store memories) $10.00 per million tokens, Recall (retrieve memories) $0.75 per million tokens, Reflect (synthesize patterns) $0.05 per call, Iris Extract (structured extraction) $7.50 per million tok
Key features
Distributed simulation across MPI ranks
Multi-backend execution (CPU, GPU, Browser, MPI)
AtlasIR typed intermediate representation
Native forward and reverse execution for training
Retains Brian2 modelling interface
GPU execution on Apple Metal and NVIDIA CUDA
WebAssembly browser execution
CPU reference and AOT compilation
Per-user persistent memory with preferences, history and decisions kept separate
Cross-session persistence allows context to survive weeks between interactions
Fast parallel retrieval in under 100ms through four strategies running simultaneously
Works with any LLM without losing learned context when swapping models
Learns from mistakes and corrects behavior on repeated problems
Automatic pattern detection synthesizes individual facts into consolidated knowledge
Conflict detection preserves both old and new states with timestamps when facts change
Integrates with 20+ coding agents including Claude Code, Cursor, and Copilot
What makes it different
Distributed execution without sacrificing Brian2's familiar modelling interface
Open-source with published arXiv paper and reproduction materials
Explicit, verifiable execution plans via AtlasIR
94.6% retrieval accuracy on LongMemEval-S, beating next-best system at 74.0%
MIT licensed open source with no usage limits, telemetry or restrictions
Reflection layer automatically builds judgment and mental models over time without manual tagging
Entity resolution handles hundreds of entities with 99.5% query reduction
Brian2 Atlas
Vectorize