Skip to content
Documentation preview. Mpalo Inc. is an early-stage Delaware Public Benefit Corporation. Palo Bloom is in development and is not yet publicly available. These pages describe intended interfaces and current development direction.

Memory Engines

Palo Bloom engines are the foundation of your memory-augmented applications. Every engine shares a core processing pipeline designed to handle episodic memory encoding, retrieval, and contextual enrichment. Selection should be based on your throughput and complexity requirements.

Engine Selection

Engines are categorized by their depth of memory processing and capacity for long-horizon recall.

Palo Nano

mpalo-palo-nano

Best for

  • Speed
  • Volume
  • Cost

Use case

  • Chatbots
  • Daily drivers
  • Batch processing

Pricing per 1M memory tokens

Encode$0.234Recall + Render$0.304

Rate limit

Defined by account plan

Latency profile

Optimized for speed

Palo Bloom

mpalo-palo

Best for

  • Balanced workloads
  • Persistent memory
  • Blurry memory

Use case

  • Assistants
  • Customer products
  • Agentic systems

Pricing per 1M memory tokens

Encode$0.486Recall + Render$0.632

Rate limit

Defined by account plan

Latency profile

Balanced for production

Palo Research

mpalo-palo-research

Best for

  • Connections
  • Knowledge
  • Long-horizon recall

Use case

  • Robotics
  • Research
  • Agentic systems

Pricing per 1M memory tokens

Encode$1.80Recall + Render$2.34

Rate limit

Defined by account plan

Latency profile

Optimized for depth

At a glance

Engine API Name Best For
Palo Nano mpalo-palo-nano High-throughput, cost-sensitive, real-time personalization; efficient for high-frequency batch operations.
Palo Bloom mpalo-palo Flagship production engine for agents and copilots.
Palo Research mpalo-palo-research Complex research agents, deep recalling, multi-session synthesis; ideal for high-latency-tolerant precision applications.

Usage rates and pricing tiers are defined in your account dashboard. Optional features like Memory Traversal and Memory Mapping are invoked via the API and billed as additional operations.

Beyond Retrieval: Memory-Aware Embeddings

Palo Bloom does more than handle conversational context. Our engines produce dense, memory-aware embeddings that embed the user's personal history and behavioral context directly into the vector representation.

For static semantic search over documents, commodity embeddings may be enough. Palo is designed for memories that need to persist, evolve, be recalled, rendered, traversed, and become more personal over time.

These embeddings are optimized for advanced workflows:

Advanced RAG

Go beyond keyword or simple semantic similarity. Retrieval can prioritize content that aligns with the user's evolving context and past history.

Personalized Recommendations

Recommendations that evolve with user intent, whether for media, content, or product discovery -- memory-biased at the embedding level.

Predictive Behavior Modeling

Use the episodic structure within the embedding to build applications that anticipate user needs based on past temporal patterns.

Memory Operations

Memory Encode

Creates Palo's latent memory embedding and rich, experience-aware memory representation.

Recall + Render

Retrieves stored memories and renders context for your external LLM.

Memory Traversal

Enables multi-hop retrieval across a user's personal timeline.

Memory Mapping

Advanced multi-path retrieval for applications requiring high-fidelity synthesis over extended histories.

Dreaming

Optional background adaptation and memory reconsolidation for products where memory quality compounds over time.

Capabilities

  • -- Input Handling: Each engine supports text, image, and (in beta) audio processing.
  • -- Storage: Operate using your own vector store or Mpalo-managed storage.
  • -- Engine Migration: As your needs grow, you may migrate memory between engine tiers via the API. This process ensures your memory is re-optimized for the target engine's capabilities.

Known Limitations

Known limitations regarding non-Latin scripts, complex visual styling, and spatial reasoning are tracked in the Project Known Issues registry. These are subjects of active iterative improvement.