mengram
Source: github.com/alibaizhanov/mengram Language: Python + JS SDKs | Status: Active (112 stars)
Human-like memory for AI agents with semantic, episodic, and procedural memory types — procedures evolve from failures.
What It Does
mengram is a cloud-hosted memory service that gives AI agents persistent, personalized memory across sessions. Its key differentiator is procedural memory — workflows that automatically evolve when they fail, creating an improvement loop.
Key Features
- 3 memory types: Semantic (facts), Episodic (events), Procedural (evolving workflows)
- Cognitive Profile — persistent user profile loaded at session start
- Claude Code hooks: auto-save after responses, auto-recall on prompts
- File upload (PDF, DOCX, TXT, MD) with vision AI extraction
- Knowledge graph
- Multi-user isolation
- Import from ChatGPT / Obsidian
- Python + JavaScript SDKs, REST API
- LangChain, CrewAI, MCP integrations
- Free tier available
Comparison to kb-mcp
| Aspect | mengram | kb-mcp |
|---|---|---|
| Hosting | Cloud (mengram.io) | Local (your machine) |
| Data control | Third-party cloud | On-disk, fully private |
| Memory model | 3-tier cognitive (semantic/episodic/procedural) | Document collections with sections |
| Search | Semantic (cloud API) | BM25 + optional local vector |
| Auto-capture | Yes (Claude Code hooks) | No (manual or agent-driven) |
| Dependencies | API key + network | Zero (Rust binary) |
Relationship: Different trust and deployment models. mengram is convenient (auto-save, cloud sync, SDKs) but sends your data to a third party. kb-mcp keeps everything local and private.
Patterns Worth Adopting
- Procedural memory — workflows that evolve from failure analysis. Relevant to the vault’s Knowledge Keeper pattern.
- Cognitive Profile — a structured “who is the user” document. Claude Code’s memory system does something similar.
- Auto-save hooks — capturing knowledge without manual intervention. The researcher agent’s heartbeat scheduling aims at this.