Skip to main content

Architecture Overview

Memorose is a memory runtime for agents. The system ingests raw events, consolidates them into durable memory, derives higher-order insights, tracks goal execution, and exposes all of that through a Rust server, a dashboard, and Raft-backed deployment primitives.

Workspace Shape​

crates/
├── memorose-common    # shared config, core types, task and graph models
├── memorose-core      # engine, storage, retrieval, graph, organization knowledge
└── memorose-server    # Axum API, dashboard auth, management routes

Runtime Pipeline​

  1. Events arrive on /v1/users/:user_id/streams/:stream_id/events
  2. L0 raw events are stored and queued for consolidation
  3. Consolidation produces L1 memory units
  4. Reflection and graph/community analysis produce L2 insights
  5. L3 goals and tasks coordinate future work and execution state
  6. Retrieval merges vector, text, graph, and shared-knowledge signals
  7. Forgetting prunes or compacts low-value memory over time

Key Product Concepts​

L0-L3​

Memorose uses an explicit memory hierarchy:

  • L0: raw event stream
  • L1: stable facts and procedural traces
  • L2: themes, clusters, reflective summaries, and shared insights
  • L3: goals, milestones, dependencies, and execution status

Domains​

The ownership and sharing model is:

  • agent: how a specific agent learns to operate
  • user: who the user is and what they prefer
  • organization: reusable shared knowledge projected from authorized source memory

Streams​

Every ingest and retrieval request is scoped to a stream_id. Streams preserve session-local chronology while still feeding long-lived memory.

Storage Model​

Memorose combines several storage engines instead of pushing everything into one system:

  • RocksDB for local durable state and key-value access
  • Lance for embeddings and vector retrieval
  • Tantivy for text retrieval
  • Graph and organization-knowledge views built on top of those primitives

Retrieval Model​

Retrieval is hybrid by design. Queries can combine:

  • semantic similarity
  • text search
  • graph depth expansion
  • time filters
  • organization knowledge
  • optional multimodal embedding input

Deployment Model​

  • Single-node mode for local development
  • Raft-based clustering for replicated deployment
  • Sharded topologies for larger installations
  • Separate dashboard UI on port 3100

This is infrastructure software first. The docs should be read with that model in mind.