DEV Community

Cover image for I built an AI memory API with 3 memory types — here's why facts alone aren't enough published
Ali Baizhanov
Ali Baizhanov

Posted on

I built an AI memory API with 3 memory types — here's why facts alone aren't enough published

#ai #javascript #python #opensource

Every AI memory tool I tried had the same problem: they only store facts.

"User likes Python." "User lives in Almaty." Cool. But human memory doesn't work like that. We have three types:

  • Semantic — facts and knowledge ("Python is a programming language")
  • Episodic — events and experiences ("I spent 3 hours debugging that auth bug last Tuesday")
  • Procedural — how to do things ("Deploy: build → upload → push → verify")

I built Mengram to give AI all three. Here's what I learned building it.

The Problem

I was building AI agents that needed to remember things across sessions. Tried Mem0, tried rolling my own with pgvector. Same issue every time:

My agent could remember that I use Railway for hosting. But it couldn't remember that last Friday's deploy broke because I forgot to run migrations. And it definitely couldn't remember that the correct deploy process is: test → build → push → migrate → verify.

That's three different kinds of memory, and every existing tool only handles the first one.

The Solution: 3 Memory Types from 1 API Call

Mengram extracts all three types automatically:

from mengram.cloud.client import CloudMemory

m = CloudMemory(api_key="om-...")

m.add([
    {"role": "user", "content": "Fixed the auth bug today. The problem was API key cache TTL was set to 0. My debug process: check Railway logs, reproduce locally, fix and deploy."}
])
Enter fullscreen mode Exit fullscreen mode

One call. Mengram's LLM extraction pipeline produces:

  • Semantic: "API key cache TTL of 0 caused auth bug"
  • Episodic: "Debugged auth bug, root cause was cache TTL, fixed and deployed"
  • Procedural: "Debug process: check logs → reproduce locally → fix → deploy"

The Killer Feature: Procedural Learning

This is what no competitor has.

Your AI agent completes a multi-step task. Mengram saves the steps as a procedure with success/failure tracking. Next time a similar task comes up, the agent already knows the optimal path.

Day 1: Agent figures out deployment
  → test → build → push → migrate → verify
  → Mengram saves as procedure (1 success, 0 failures)

Day 5: Agent deploys again  
  → Finds procedure in memory
  → Follows proven path
  → Records success (2 successes, 0 failures)

Day 12: Agent skips tests, deploy breaks
  → Records failure (2 successes, 1 failure)
  → Next time: "This procedure works better with tests first"
Enter fullscreen mode Exit fullscreen mode

The AI literally learns from its own experience. Not from fine-tuning, not from few-shot examples — from actual procedural memory.

Smart Triggers: Memory That Raises Its Hand

Most memory is passive — you ask, it answers. Mengram also has proactive memory:

  • Reminders: "You mentioned a meeting with Anya tomorrow at 3pm" → fires webhook 1 hour before
  • Contradictions: Memory says "Anya is vegetarian" → you say "order steaks for dinner with Anya" → alert
  • Patterns: 3 out of 5 Friday deploys had bugs → "Maybe wait until Monday?"

These fire automatically via webhooks — works with Slack, Discord, OpenClaw, or any endpoint.

Integrations

Mengram works as a memory layer for any AI stack:

  • Claude Desktop — MCP server, just add to config
  • LangChain — drop-in MengramMemory class replacing ConversationBufferMemory
  • CrewAI — 5 tools including mengram_save_workflow for procedural learning
  • OpenClaw — skill on ClawHub with bash scripts for all channels
  • Any LLM — REST API + Python/JS SDKs 
pip install mengram-ai    # Python
npm install mengram-ai    # JavaScript
Enter fullscreen mode Exit fullscreen mode

Cognitive Profile

One API call generates a system prompt from everything Mengram knows about a user:

profile = m.get_profile()
print(profile["system_prompt"])
Enter fullscreen mode Exit fullscreen mode 
You are talking to Ali, a 22-year-old developer in Almaty building Mengram.
He uses Python, PostgreSQL, and Railway. Recently: debugged pgvector deployment,
researched competitors. Workflows: deploys via build→twine→npm→git.
Communicate in Russian/English, direct style, focus on practical next steps.
Enter fullscreen mode Exit fullscreen mode

Insert into any LLM for instant personalization. Replaces your RAG pipeline.

Architecture

Built on PostgreSQL + pgvector. No separate vector database needed.

Your AI Client (Claude, GPT, any LLM)
    │
    ▼
Mengram Cloud API
    ├── LLM Extraction (entities, episodes, procedures)
    ├── Embedding (OpenAI text-embedding-3-large)
    ├── Hybrid Search (vector + full-text + re-ranking)
    ├── Smart Triggers (reminders, contradictions, patterns)
    └── Memory Agents (Curator, Connector, Digest)
    │
    ▼
PostgreSQL + pgvector
    ├── Entities & Facts (semantic)
    ├── Episodes (episodic)
    ├── Procedures (procedural)
    └── Embeddings (1536-dim vectors)
Enter fullscreen mode Exit fullscreen mode

What I Learned

1. Extraction is everything. The quality of your memory system depends entirely on how well you extract structured data from conversations. I went through 3 versions of the extraction prompt before it reliably separated facts from events from procedures.

2. Contradiction detection is harder than it sounds. "I'm vegetarian" and "I love steak" — obvious contradiction. "I prefer dark mode" and "I switched to light mode" — is that a contradiction or an update? LLM-based conflict resolution was the answer.

3. Procedural memory is the moat. Every competitor does semantic memory. Some do episodic. Nobody does procedural with success/failure tracking. This is what makes agents genuinely learn from experience.

Try It

Free tier, no credit card, 60-second setup:

  1. Sign up at mengram.io
  2. pip install mengram-ai
  3. Start adding memories

Open source (Apache 2.0): github.com/AiBaizhanov/mengram

API docs: mengram.io/docs

I'd love feedback — especially from anyone building AI agents. What memory challenges are you running into?

Top comments (1)

Collapse
 
alibaizhanov profile image
Ali Baizhanov •

Hey Product Hunt! I'm Ali, a 32-year-old developer from Almaty, Kazakhstan.
I built Mengram because every AI memory tool I tried only stored facts. But human memory has 3 types — we remember facts (semantic), events (episodic), and how to do things (procedural). So I built an API that does all three.
The killer feature: your AI agent completes a task → Mengram saves the steps as a procedure → next time, it already knows the optimal path. No other memory API does this.
It's free, open-source, and takes 60 seconds to set up. Would love your feedback!