Human-like memory layer for AI agents featuring semantic, episodic, and procedural memory with failure-driven workflow evolution. Offers Python/JS SDKs, REST API, and LangChain/CrewAI integrations.
Project Positioning#
Mengram provides human-like memory capabilities for AI agents, addressing the core problem of AI agents lacking long-term memory and context retention. Agents cannot automatically optimize workflows from past failures or successes. Mengram offers a unified memory layer connecting different agent frameworks.
Core Memory Architecture#
The project adopts a three-layer memory model:
- Semantic Memory: Stores facts, preferences, knowledge; retrieved via
m.search("tech stack")returning fact lists - Episodic Memory: Stores events, decisions, outcomes; queried via
m.episodes(query="deployment")returning events with results and dates - Procedural Memory: Evolvable workflows with version history; accessed via
m.procedures(query="deploy")returning workflow version history
Intelligent Evolution Mechanism#
Trigger automatic workflow evolution via m.procedure_feedback(proc_id, success=False, context="..."). The system automatically detects and evolves workflows from failed conversations, achieving experience-driven continuous optimization. Key data flow: POST /v1/add → session extraction → store entities/events/procedures → auto-associate failure events and evolve procedures.
Retrieval & Cognitive Capabilities#
- Unified Retrieval:
m.search_all("deployment issues")searches across all three memory types - Cognitive Profile:
m.get_profile()generates personalized summaries usable directly as system prompts - Knowledge Graph: Supports graph-based storage and retrieval of entities, relationships, and facts
Data Management#
- Multi-user Isolation: Data isolation via
user_id(sub_user_id) - Data Import: CLI supports
mengram import chatgpt export.zip --cloud,mengram import obsidian vault --cloud,mengram import files notes/*.md --cloud - Smart Triggers: Reminders, contradiction detection, pattern recognition
- Webhooks: Change event notifications
Ecosystem Integrations#
- MCP Server: Supports Claude Desktop, Cursor, Windsurf configuration
- LangChain: Provides
MengramChatMessageHistory,MengramRetriever - CrewAI: 5 dedicated memory tools
- OpenClaw: Plugin supports automatic recall and capture with 12 tools
Deployment Modes#
- Cloud Mode: FastAPI + PostgreSQL + pgvector, includes session extractor, evolution engine, smart triggers, memory agent
- Local Mode: MengramBrain engine, SQLite + vector indexing + knowledge graph, hybrid retrieval without external database dependencies
Quick Start#
from cloud.client import CloudMemory
m = CloudMemory(api_key="om-...") # Get free key from mengram.io
m.add([{"role": "user", "content": "I use Python and deploy to Railway"}])
m.search("tech stack") # → list of facts
m.episodes(query="deployment") # → list of events
m.procedures(query="deploy") # → workflow version history
Use Cases#
- DevOps Agent: Automatically evolves deployment workflows from failure conversations (example:
examples/devops-agent) - Customer Support: Remembers returning customer preferences and history, combined with CrewAI 5 memory tools
- Personal Assistant: Builds assistants with personalized cognitive profiles using LangChain (example:
examples/personal-assistant)