mengram-ai 1.0.5

AI memory layer for apps. Open-source Mem0 alternative. Like Mem0, but you own your data.

🧠 Mengram

AI memory layer for apps. Open-source Mem0 alternative.

Build persistent memory for AI agents and apps. Knowledge graph, semantic search, Cloud API & MCP server. Use locally with Obsidian vault or via cloud.

Like Mem0, but you own your data — and it actually saves your solutions with code, not just "user uses PostgreSQL".

---

Why Mengram?

	Mem0	Basic Memory	Mengram
Storage	Cloud vectors	Flat markdown	Typed knowledge graph in .md
Entity types	❌ Flat facts	❌ One note per chat	✅ Person, Project, Technology, Company
Relations	❌	❌	✅ works_at, uses, depends_on
Rich knowledge	❌	❌	✅ Solutions, configs, formulas with code
Proactive context	❌	❌	✅ Auto-injected — no manual recall
Obsidian graph	❌	Partial	✅ Full [[wikilinks]] + graph view
Semantic search	✅ Cloud	❌	✅ Local embeddings (384D)
Own your data	❌ Cloud lock-in	✅	✅ Plain .md files
LLM agnostic	❌	Partial	✅ Claude / GPT / Ollama
Pricing	$24/mo+	$14/mo	Free & open source

What it actually does

You chat with Claude (or any LLM). Mengram automatically:

1. Extracts entities, facts, relationships, and rich knowledge (solutions, commands, configs with code)
2. Creates typed .md files in your Obsidian vault
3. Links everything with [[wikilinks]] and YAML frontmatter
4. Indexes with local vector embeddings for semantic search
5. Proactively injects relevant context into every conversation — no manual recall needed

You: "We fixed the OOM with Redis cache. Config: hikari.pool-size=20"

         ┌─────────────────────────────────────┐
         │  vault/PostgreSQL.md                 │
         │  type: technology                    │
         │                                      │
         │  ## Facts                             │
         │  - Main database, version 15          │
         │                                      │
         │  ## Knowledge                         │
         │  **[solution] Connection pool fix**    │
         │  OOM at 200+ WebSocket → Redis cache  │
         │  ```yaml                              │
         │  spring.datasource.hikari.            │
         │    maximum-pool-size: 20              │
         │  ```                                  │
         └─────────────────────────────────────┘

Next time you ask "How did we fix the OOM?" → Claude already knows, with the config.

---

Quick Start

1. Install

pip install mengram-ai[all]

2. Setup (one command)

mengram init

This will:

• Ask for your LLM provider and API key
• Create ~/.mengram/config.yaml and vault
• Auto-configure Claude Desktop MCP integration
• Tell you to restart Claude Desktop

That's it. Talk to Claude — it remembers automatically and always has context.

Non-interactive:

mengram init --provider anthropic --api-key sk-ant-...

Other commands:

mengram status    # Check setup
mengram stats     # Vault statistics
mengram server    # Start MCP server manually

---

Proactive Context (v0.5.0)

The killer feature. Claude Desktop gets your knowledge profile automatically — no manual attach, no "recall", no "remember what I told you".

How it works:

Claude Desktop starts
  → MCP server reads vault
  → Generates compact knowledge index (scales to 1000+ notes)
  → Injects into Claude's instructions
  → Warms up semantic search model

You open any chat → Claude already knows:
  - Your tech stack, projects, team
  - Past solutions with code/configs
  - Entity relationships

You ask a question → Claude auto-calls recall()
  → Gets full details + code artifacts
  → Answers with context

---

Rich Knowledge (v0.5.0)

Not just "user uses PostgreSQL" — but solutions with code, commands, formulas, configs.

The LLM automatically chooses the knowledge type based on context:

Domain	Knowledge types	Example
Developer	solution, command, config, debug	HikariCP pool config with YAML
Doctor	treatment, lab_result, diagnosis	Metformin 500mg dosage
Scientist	experiment, formula, hypothesis	Protein denaturation at 60°C
Student	formula, example, insight	Bayes theorem with example
Chef	recipe, tip, substitution	Sourdough hydration ratio

No configuration needed. The system adapts to any domain.

## Knowledge

**[solution] Connection pool exhaustion fix** (2024-02-10)
OOM at 200+ WebSocket connections → Redis cache for UserService
​```yaml
spring.datasource.hikari.maximum-pool-size: 20
spring.datasource.hikari.idle-timeout: 30000
​```

**[command] Debug database connections** (2024-02-10)
Monitor active PostgreSQL connections
​```sql
SELECT count(*), state FROM pg_stat_activity GROUP BY state;
​```

---

Python SDK (Mem0-compatible API)

from mengram import Memory

m = Memory(
    vault_path="./my-brain",
    llm_provider="anthropic",
    api_key="sk-ant-..."
)

# Remember — extracts entities, facts, relations, AND knowledge
m.add("I work at Uzum Bank, backend on Spring Boot and PostgreSQL", user_id="ali")

# Semantic search (finds by MEANING, not just keywords)
results = m.search("database issues", user_id="ali")
for r in results:
    print(f"{r.memory.name} (score={r.score:.2f})")
    print(r.memory.facts)

# Get everything
all_memories = m.get_all(user_id="ali")

# Stats
print(m.stats(user_id="ali"))

Auto-Memory Middleware

Drop-in wrapper that automatically remembers and recalls:

from mengram import Memory
from mengram_middleware import AutoMemory

m = Memory(vault_path="./vault", llm_provider="anthropic", api_key="sk-ant-...")
auto = AutoMemory(memory=m, user_id="ali")

# Automatically: recall context → inject → LLM response → remember new knowledge
response = auto.chat("Help me fix the PostgreSQL connection pool issue")

---

How It Works

Conversation → Extractor (LLM) → Entities + Facts + Relations + Knowledge
                                          ↓
                                   Vault Manager → .md files with [[wikilinks]]
                                          ↓
                                   Vector Index → local embeddings (SQLite)
                                          ↓
                                   MCP Server → instructions (compact index)
                                              → tools (recall, remember)
                                          ↓
                                   Claude Desktop → auto-context every chat

Semantic Search (Hybrid)

3-level search strategy:

1. Vector Search — all-MiniLM-L6-v2 (80MB, runs locally). Finds "database" when you search "PostgreSQL" — by meaning, not keywords.
2. Graph Expansion — follows [[wikilinks]] from top results. Found PostgreSQL? Also returns linked Project Alpha.
3. Text Fallback — substring match for edge cases.

Entity Types

Type	Examples
person	Team members, contacts
project	Services, repos, products
technology	PostgreSQL, Spring Boot, Kafka
company	Employers, clients, partners
concept	Patterns, strategies, ideas

File Format

---
type: technology
created: 2024-02-10 15:30
updated: 2024-02-11 09:15
tags: [technology]
---

# PostgreSQL

## Facts

- Main database, version 15
- Connection pool issue in [[Project Alpha]]

## Relations

- ← uses [[Project Alpha]]: Main DB

## Knowledge

**[solution] Connection pool exhaustion fix** (2024-02-10)
OOM at 200+ WebSocket → Redis cache for UserService
​```yaml
spring.datasource.hikari.maximum-pool-size: 20
​```

---

Configuration

# config.yaml
vault_path: "./vault"

llm:
  provider: "anthropic"  # anthropic | openai | ollama | mock
  anthropic:
    api_key: "sk-ant-..."
    model: "claude-sonnet-4-20250514"

semantic_search:
  enabled: true

Provider	Install	Cost
Anthropic (Claude)	pip install mengram-ai[anthropic]	API pricing
OpenAI (GPT)	pip install mengram-ai[openai]	API pricing
Ollama (local)	Install ollama	Free

---

Roadmap

• Typed entity extraction (person, project, technology, company)
• Obsidian vault with [[wikilinks]] + YAML frontmatter
• MCP Server for Claude Desktop
• Semantic search with local embeddings
• Hybrid retrieval (vector + graph)
• Mem0-compatible Python SDK
• Auto-memory middleware
• Rich knowledge extraction (solutions, configs, formulas with code)
• Proactive context (auto-injected via MCP instructions)
• Entity deduplication
• Obsidian plugin (TypeScript)
• Web dashboard
• REST API

Contributing

git clone https://github.com/alibaizhanov/mengram
cd mengram
pip install -e ".[all,dev]"
pytest

License

MIT