Knowledge Graph
Link, traverse, and query — no external graph DB.
[ STRUCTURED MEMORY SDK FOR AI AGENTS ]
Memory that agents can actually reason over.
V0.9.1 — NOW AVAILABLEHow It Works
Memory that follows your users.
No external services. Everything runs on your machine.
Agent Chat
I prefer aisle seats when flying. But for long flights, make that a window seat.
Noted. Aisle seats by default, window seats for long flights — all set.
2 weeks later
Can you book my flight to Berlin next Tuesday? It’s a long one.
Booking it now — window seat as per your preferences.
Memory Store
No memories yet
PREFERENCE1.00
Prefers aisle seats for flights
PREFERENCE1.00
Prefers window seat for long flights
PREFERENCErecalled
Prefers window seat for long flights
[ HOW MEMORY GETS STRUCTURED ]
From conversation to knowledge.
MemBlock extracts structured memory blocks from raw conversations — typed, scored, and ready for retrieval.
User said:
“I prefer Python, work at Acme Corp since 2024, and deploy on AWS.”
PREFERENCE
Prefers Python
FACT
Works at Acme Corp
EVENT
Joined Acme Corp
Core Capabilities
Everything an agent needs to remember
Every primitive your AI agent needs for structured, searchable, and secure long-term memory.
Memory Decay Engine
Prune weak memories, surface the strongest.
Temporal Reasoning
Time-aware retrieval with natural language time parsing.
Multi-Hop Retrieval
Entity extraction + graph walk for complex reasoning.
Adaptive Context
Graph-expanded, deduplicated context with confidence gating.
Async + Event Hooks
Full async API with lifecycle event hooks.
Production Ready
Ship to production, not just a demo
Everything you need to deploy AI memory at scale — from multi-tenant isolation to LLM-powered extraction and tamper-proof audit trails.
PostgreSQL + pgvector
Production-grade storage with full-text search and native vector search built in.
Multi-Tenant Isolation
User-scoped data at every level — org, project, session, and agent.
Connection Pooling
Efficient pooling with per-user instances for high-concurrency workloads.
LLM Auto-Extraction
Extract structured memories from conversations using any major LLM provider.
Org-Level Analytics
Track questions across users with trending topics and noise filtering.
Per-Block Encryption
AES-256-GCM encryption at the block level — encrypt what matters, leave the rest fast.
Use Cases
Long-term memory for any scenario
AI Companion
Persistent user context across every session.
Coding Agents
Remember codebase patterns and developer preferences.
Customer Support Bot
Track tickets, history, and resolution patterns.
Education Agent
Adapt to learning pace and knowledge gaps.
Gaming & Interactive
NPCs that remember player choices and relationships.
Creation Assistants
Recall style guides, past edits, and creative context.
Head-to-Head Benchmarks
97.2% Recall@5 on LongMemEval
MemBlock beats MemPalace (96.6%) on LongMemEval with zero API calls. Tested head-to-head on the same datasets, same metrics, same machine. 10x faster retrieval with SQLite vs ChromaDB.
VIEW FULL RESULTS97.2%LongMemEval R@5
57.7%LoCoMo R@5
LocalZero API Calls
10xvs MemPalace Speed
Quickstart
Get started in three lines of Python.
Install from PyPI. Create a MemBlock instance. Store your first memory. No API keys, no configuration files, no setup wizard.
VIEW DOCUMENTATIONpython3
$ pip install memblock# Store structured memories>>> from memblock import MemBlock, BlockType>>> mem = MemBlock(storage="sqlite:///memory.db")>>> mem.store("User prefers dark mode", type=BlockType.PREFERENCE)Block(id='blk_7f2a', type=PREFERENCE, confidence=1.0)# Query by text>>> mem.query(text_search="dark mode")[Block(id='blk_7f2a', content='User prefers dark mode')]# Build LLM-ready context>>> mem.build_context(query="user preferences", token_budget=2000)'[PREFERENCE] User prefers dark mode (confidence: 1.00)'