MemTensor/MemOS
MemOS
MemOS is an open-source persistent memory layer for AI agents and large language models, providing a self-hosted server that stores and retrieves structured memory across sessions. It enables AI systems to recall user preferences, history, and context without retraining, using a graph-based API and a web management interface for viewing, editing, and organizing memory items, skills, traces, and knowledge bases.
The system distinguishes itself through a portable memory interchange protocol that allows memory to be transferred between different AI models, devices, and applications, along with a three-tier memory architecture that organizes information into Skills, Traces/Episodes, and World Models. It stores persistent memory as human-readable Markdown files on local SQLite storage rather than opaque vector databases, and supports hybrid semantic-lexical retrieval that combines vector cosine similarity with BM25 lexical search. Additional differentiators include predictive memory preloading that loads relevant context before it is needed, asynchronous memory ingestion for millisecond-level latency under high concurrency, and memory-based skill crystallization that extracts repeated strategies into callable, versioned skills.
The platform offers composable knowledge base cubes for managing multiple isolated knowledge bases with controlled sharing across users, projects, and agents, along with per-agent memory isolation and feedback-based memory refinement. It supports multi-modal memory storage for text, images, tool traces, and personas, and provides a CLI tool for cross-agent memory sharing. Memory lifecycle management includes full CRUD operations, batch cleanup, tagging, and governance, while retrieval budget control limits token consumption by capping the number of memories recalled per task. The system also enables team skill sharing over LAN or VPN and supports data import and export in JSON, legacy plugin, and agent-specific formats.
Documentation covers deployment across public cloud, private cloud, on-premises, and hybrid architectures, with the server running on SQLite for privacy and offline operation.
Features
- AI Memory Layers - Provides a persistent memory layer for AI agents to recall user preferences and history across sessions.
- Graph-Structured Memory APIs - Adds, retrieves, edits, and deletes long-term memory through a single graph-structured API that remains inspectable and editable.
- Long-term Memory Injection - Integrates persistent memory into AI agents with a few lines of code, enabling recall across sessions and tasks.
- Composable Knowledge Cubes - Manages multiple isolated knowledge bases as modular cubes with controlled sharing.
- Local SQLite Stores - Stores all AI memory locally using SQLite, ensuring privacy and offline operation.
- Agent Memory Stores - Provides persistent storage for maintaining state, user preferences, and conversation history across agent sessions.
- Hybrid Search Retrievers - Combines vector cosine similarity with BM25 lexical search for hybrid memory retrieval.
- Three-Tier Memory Architectures - Organizes memory into three tiers with separate retrieval channels for Skills, Traces, and World Models.
- AI Agent Servers - Provides an open-source server that stores and retrieves structured memory for AI agents across sessions.
- Context-Aware Retrieval - Searches stored memories by user and query, returning relevant results to personalize AI agent interactions.
- Conversation Memory Stores - Ships a conversation memory store that persists and retrieves interaction history for AI agents.
- Memory Persistence - Provides memory persistence so AI agents recall past interactions and user preferences across sessions.
- Memory Portability Utilities - Transfers persistent memory between AI models, devices, and applications via a portable protocol.
- Millisecond Retrieval Systems - Returns relevant memory records with millisecond latency, enabling real-time recall during AI interactions.
- Markdown Memory Stores - Stores persistent AI memory as human-readable Markdown files instead of opaque databases.
- Memory Ingestion Pipelines - Processes memory operations asynchronously with millisecond latency for high concurrency.
- SQLite Storage Adapters - Keeps all memory data and logs on the user's machine using SQLite for privacy and offline operation.
- Memory Entry Lifecycles - Provides full CRUD, batch cleanup, tagging, and governance for AI memory entries.
- Cross-Model Memory Protocols - Shares persistent memory between different AI models via a portable interchange protocol.
- Cross-Model Memory Protocols - Transfers persistent memory between different AI models using a portable interchange protocol.
- Cross-Model Memory Sharing - Shares persistent memory between different models, devices, and applications using a portable memory interchange protocol.
- Experience-Based Skill Extractors - Extracts repeated high-value strategies into callable, versioned skills that improve with feedback.
- Feedback Loops - Corrects, supplements, or replaces existing memories using natural-language feedback to improve accuracy over time.
- Shared Skill Distribution - Enables optional sharing of crystallized skills and trace excerpts over LAN or VPN through a dedicated panel.
- Low-Latency Retrieval Systems - Returns stored memory records with millisecond latency for real-time AI recall.
- Prompt Assembly Systems - Formats retrieved memory fragments into a structured prompt with safety rules to guide an AI's personalized response.
- Memory Retrieval Budget Controllers - Limits the number of memories recalled per task to reduce token consumption and avoid irrelevant context.
- Hybrid Search Methods - Combines vector cosine similarity with BM25 lexical search, fusing scores to rank relevant memories.
- Multi-Modal Memory Stores - Stores and retrieves text, images, tool traces, and personas together for richer AI context.
- Management Interfaces - Ships a web UI for managing AI memory items, skills, traces, and knowledge bases.
- Tool Output Compressors - Provides compression and deduplication of long tool outputs to maintain clean AI memory.
- Asynchronous Ingestion Pipelines - Processes memory operations asynchronously with millisecond-level latency under high concurrency.
- Multi-Modal Stores - Stores and retrieves text, images, tool traces, and personas together in one memory system.
- Natural Language Memory Queries - Queries stored memories by a natural-language prompt and returns matching factual and preference records.
- Knowledge Base Management - Organizes memory into knowledge bases with configurable storage per item and a maximum number of bases per plan.
- Composable Cubes - Manages multiple isolated knowledge bases as composable memory cubes for controlled sharing across users and agents.
- Multi-Modal Memory Stores - Stores and retrieves text, images, tool traces, and personas together in one memory system.
- Three-Tier Retrieval Systems - Searches across Skills, Traces/Episodes, and World Models using multiple retrieval channels with RRF and MMR.
- API-Based Stores - Stores persistent memory entries via an API with monthly quotas ranging from 50K to unlimited entries.
- Experience-Based Skill Extractors - Extracts repeated strategies into callable, versioned skills that improve with feedback.
- Self-Hosted Infrastructure - Runs on public cloud, private cloud, on-premises, or hybrid architectures using SQLite for full data control.
- Memory Isolation - Separates memory storage and retrieval by agent identifier with optional whitelist control and per-agent configuration.
- Predictive Memory Preloading - Loads relevant memory before it is needed by analyzing dialogue history, task semantics, or environmental cues.
- Shared Filesystem Memory - Transfers persistent memory between different AI models, devices, and applications through a portable memory protocol.
- Memory Management Panels - Ships a web management panel for viewing, editing, and organizing AI memory items and skills.
- AI - Provides full CRUD, batch cleanup, tagging, and governance operations for controlling memory throughout its lifecycle.
- Memory - Searches stored memory entries via an API with monthly quotas ranging from 20K to unlimited searches.
