SurrealDB vs. memory middleware
Mem0, Zep, Letta, and other memory middleware add another layer between your agent and storage. The fragmented memory tax is why memory should live in the database.
Memory middleware sits between the agent and the database, introducing consistency gaps, added latency, separate permissions, and operational burden. Every memory operation becomes a round trip through another system with its own failure modes.Spectron eliminates the middleware layer entirely - memory lives in the database.
THE LANDSCAPE
A growing category with a shared limitation
A growing number of agent memory middleware products have emerged in the last year. Each product solves a real developer pain point, but they all share the same architectural limitation: memory lives outside the database, not inside it.
Mem0
LLM-driven memory extraction and consolidation with optional graph storage. Excellent developer experience, but built on fragmented infrastructure underneath.
Zep
Long-term memory for assistants with temporal knowledge graphs and entity extraction. Adds structure, but as a separate service with its own consistency model.
Letta (MemGPT)
Self-editing memory framework where agents manage their own context window and long-term storage. Innovative, but memory lives outside the transactional boundary.
Cognee
Open-source framework for building knowledge graphs and retrieval pipelines from unstructured data. Graph construction is valuable, but disconnected from application state.
LangMem
LangChain library for persisting and retrieving agent memories within LangGraph workflows. Tightly coupled to one framework with no standalone persistence guarantees.
Memori & Synix
Enterprise memory fabrics with classification, semantic search, and provenance tracking. Rich features, but still middleware between the agent and the database.
THE FRAGMENTED MEMORY TAX
Middleware hides the problem. It doesn't solve it.
Data fragments at every system boundary
Consistency gaps
Memory middleware sits between the agent and the database. Writes to memory and writes to state can diverge - no atomicity across systems.
Latency
Every memory operation is a round trip through middleware. Read from memory, then read from DB - two hops instead of one.
Operational burden
Another service to deploy, monitor, and secure. Another failure mode. Another set of credentials to manage.
Separate permissions
Memory middleware has its own auth. Your database has its own. Two permission models to keep in sync.
committed
inconsistent state
rolled back
Vector DB
Similarity search
Graph DB
Relationship store
Doc store
Raw document index
Auth service
Identity & tokens
Five systems. Five auth configs. Five things to scale, secure, and keep alive.
AGENT FRAMEWORKS
Framework memory is not database memory
LangChain, LlamaIndex, CrewAI, AutoGen, and Semantic Kernel all include memory modules. These are useful for prototyping, but they delegate persistence to external stores and offer no transactional guarantees across memory and application state.
Framework memory
Memory modules in agent frameworks provide convenience abstractions, but persistence depends on whatever backend is plugged in. No ACID, no unified permissions, no temporal awareness.
Database-native memory
Spectron provides the same memory capabilities - entity extraction, knowledge graphs, temporal facts - but as a first-class database feature with ACID transactions and unified governance.
THE ALTERNATIVE
ACID transactions across every data model
ACID memory
Spectron memory lives in SurrealDB. Agent memory and application state commit in one transaction - no divergence.
One permission model
RBAC and record-level permissions apply to memory and data alike. One auth, one policy.
No middleware
Agents write directly to the database. One round trip. No extra latency, no extra failure modes.
Vertical stack
SurrealDB + Spectron. Structured context and persistent memory in one deployment. One operational surface.
GET STARTED
Memory without the memory tax
Spectron eliminates the fragmented memory tax entirely. No middleware, no consistency gaps, no extra infrastructure.
