GitHub Repository

Human-like memory for AI agents — semantic, episodic & procedural. Experience-driven procedures that learn from failures. Free API, Python & JS SDKs, LangChain, CrewAI & OpenClaw integrations.

173 starsPython

Mengram – AI agent memory with facts, events, and evolving workflows

Name: Mengram – AI agent memory with facts, events, and evolving workflows
Availability: InStock
Author: mengram-ai

by mengram-ai·Feb 25, 2026·2 points·1 comment

Visit Project View on HN

AI Analysis

●●●BangerZero to OneBig Brain

Procedures that learn from failures—not just facts; agents improve through experience loops.

Strengths

•Novel memory model: three-tier architecture (semantic + episodic + procedural) with failure-driven procedure evolution is genuinely different from Mem0/Zep/Letta.
•Low friction: free tier, multi-SDK support (Python, JS), LangChain and CrewAI integrations, import from ChatGPT/Obsidian all present.
•Cognitive profiles and experience-weighted procedure updates address real agent drift—the 'same mistake twice' problem is well-motivated.

Weaknesses

•API-dependent architecture (no self-hosted free tier option shown); requires external service for persistence.
•Early-stage product (54 stars, limited adoption data); evolution mechanism relies on agent correctly identifying failures—brittleness not proven at scale.

Post Description

Hi HN, I built Mengram because every AI memory tool I tried only stored facts. My agents kept making the same mistakes — forgetting what happened, losing workflows.

Mengram stores 3 types: semantic (facts), episodic (events/decisions), and procedural (workflows). The key difference: procedures evolve when they fail. Week 1: deploy → build → push (fails: forgot migrations) Week 2: deploy v2 → build → migrate → push (fails: OOM) Week 3: deploy v3 → build → migrate → check memory → push This happens automatically from conversations — report a failure, the procedure evolves. Stack: Python, PostgreSQL + pgvector, FastAPI. Free cloud API, self-hostable, SDKs for Python/JS, integrations with LangChain, CrewAI, MCP. Honest limitations: extraction quality depends on LLM, procedural evolution needs clear failure descriptions, no real-time streaming yet.

Would love feedback on the memory model — is 3 types the right abstraction, or too complex?