How memtensor/memos works
MemOS is a memory system for LLM agents, offered as a hosted API, self-hosted service, and agent plugins. The main workflow accepts agent or application memory operations, searches or updates memory, and returns context; asynchronous ingestion and memory evolution add background stages. The graph also preserves the documented multi-cube, feedback, multimodal, and skill capabilities. Some subsystem links reflect README descriptions and the file tree rather than inspected implementations, so uncertain internal wiring is omitted.
Main entry points: search_handler.py Memory API · bridge.mts Local plugins · index.ts Memory storage · capture.ts Experience capture
Components 21
Interfaces
- Memory API
src/memos/api/handlers/search_handler.py - Cloud plugin
apps/MemOS-Cloud-OpenClaw-Plugin/index.js - Local plugins
apps/memos-local-plugin/bridge.mts - OpenClaw local
apps/memos-local-openclaw/src/index.ts - Memory viewer
apps/memos-local-plugin/viewer/src/components/App.tsx
Memory lifecycle
- Memory ingestion
src/memos/mem_scheduler/task_schedule_modules/handlers/query_handler.py - Memory search
src/memos/mem_scheduler/memory_manage_modules/search_pipeline.py - Local recall
apps/memos-local-plugin/core/retrieval/retrieve.ts - Experience capture
apps/memos-local-plugin/core/capture/capture.ts - Sessions and episodes
apps/memos-local-plugin/core/session/manager.ts
Learning and sharing
- Memory enhancement
src/memos/mem_scheduler/memory_manage_modules/enhancement_pipeline.py - Feedback learning
apps/memos-local-plugin/core/feedback/feedback.ts - Evolving skills
apps/memos-local-plugin/core/skill/skill.ts - World models
apps/memos-local-plugin/core/memory/l3/l3.ts
Service and state
- Memory storage
apps/memos-local-plugin/core/storage/index.ts - Users and cubes
src/memos/mem_user/mysql_user_manager.py - Async scheduler
src/memos/mem_scheduler/task_schedule_modules/handlers/query_handler.py - Memory hub
apps/memos-local-plugin/core/hub/server.ts
Other components
- Applications
- AI agents
- LLM providers
Connections 25
- Applications → Memory API: memory requests
- Memory API → Memory search: searches memories
- Memory search → Memory API: returns results
- Memory API → Users and cubes: checks user and cube
- AI agents → Cloud plugin: uses hosted memory
- Cloud plugin → Memory API: calls cloud API
- AI agents → Local plugins: uses local memory
- AI agents → OpenClaw local: uses local memory
- Local plugins → Sessions and episodes: manages episodes
- Sessions and episodes → Memory storage: persists session state
- Local plugins → Experience capture: captures experience
- Experience capture → Memory storage: stores memories
- Local plugins → Local recall: retrieves context
- Local recall → Memory storage: searches memory
- Memory API → Async scheduler: submits async work
- Async scheduler → Memory ingestion: dispatches messages
- Memory ingestion → Memory enhancement: enhances memories
- AI agents → Feedback learning: corrects memory
- Feedback learning → Memory storage: updates memory
- Experience capture → Evolving skills: supports skill evolution
- Experience capture → World models: supports abstraction
- Local plugins → Memory hub: shares memory
- Memory viewer → Memory storage: inspects memory
- Memory enhancement → LLM providers: prompts enhancement
- Local recall → LLM providers: judges relevance