SuperLocalMemory Logo — Local AI Memory Layer
SuperLocalMemory
ARCHITECTURAL COMPARISON

SuperLocalMemory
vs MemGPT / Letta

A technical comparison of two different approaches to AI agent memory: a standalone mathematical memory layer vs an OS-inspired agent framework with LLM-managed memory.

MemGPT was rebranded as Letta in 2024. Both names refer to the same project.

Important Framing

SuperLocalMemory and Letta solve different problems and are largely complementary, not competing.

SuperLocalMemory

A memory layer (backend) — provides persistent, local-first, mathematically-retrieved memories to any agent framework via MCP. Not a full agent framework.

Letta / MemGPT

A complete agent framework — the LLM itself manages memory (what to store, retrieve, summarize). Includes execution environment, not just a memory backend.

At a Glance

SuperLocalMemory V3

  • —Memory layer only — adds memory to any agent via MCP
  • —Mathematical retrieval — Fisher-Rao, sheaf, Langevin
  • —74.8% LoCoMo (zero cloud) / 87.7% full power
  • —No LLM needed for retrieval (Mode A)
  • —EU AI Act compliant by architecture (Mode A)
  • —AGPL v3 — fully open source

Letta (MemGPT)

  • —Full agent framework with LLM-managed memory
  • —OS-inspired: core context, recall, archival memory tiers
  • —~83.2% LoCoMo (cloud LLM required)
  • —LLM decides memory operations (store, retrieve, summarize)
  • —Supports cloud and local LLMs (Ollama)
  • —Apache 2.0 license — open source

Memory Architecture Approaches

Dimension SuperLocalMemory V3 Letta / MemGPT
What it is Memory layer (backend) — plug into any agent via MCP Complete agent framework with memory management built in
Memory Management Mathematical: Fisher-Rao retrieval, sheaf consistency, Langevin lifecycle LLM-driven: the model decides what to store/retrieve/summarize
Data Locality On-device (Mode A/B) — zero cloud (Mode A) Local LLM option (Ollama) or cloud LLMs; data may be local or cloud
LoCoMo Score 74.8% (Mode A zero-cloud) / 87.7% (Mode C) ~83.2% (requires LLM)
LLM Dependency None for retrieval (Mode A) — optional for synthesis (Mode C) Required — LLM controls all memory operations
EU AI Act (Mode A) Compliant by architecture — zero data transit Depends on LLM choice (cloud = compliance work required)
Integration MCP interface — works with any AI tool or agent Framework-native agents built within Letta
Transparency Every retrieval decision auditable (6-channel scores visible) LLM memory decisions — inherits LLM opacity

Benchmark Context

LoCoMo benchmark results for context. Full field comparison at superlocalmemory.com/landscape.

System LoCoMo Score Cloud LLM Required
SLM V3 Mode C 87.7% Yes (synthesis only)
Letta / MemGPT ~83.2% Yes
SLM V3 Mode A Retrieval 74.8% No
SLM V3 Mode A Raw (zero-LLM) 60.4% No

SLM V3 results from: arXiv:2603.14588.

Frequently Asked Questions

What is the difference between MemGPT/Letta and SuperLocalMemory architecturally?

+
MemGPT (rebranded as Letta) is an OS-inspired agent framework where memory management is handled by the LLM itself — the model decides what to move between context window, recall storage, and archival storage. SuperLocalMemory is a standalone memory layer that any agent can use via MCP: retrieval is mathematical (Fisher-Rao, sheaf, Langevin), not LLM-driven. SLM works as a memory backend for any agent framework including Letta.

Is SuperLocalMemory compatible with Letta?

+
SuperLocalMemory provides a standard MCP (Model Context Protocol) interface that agent frameworks can use for memory operations. Letta agents can use SuperLocalMemory as an external memory source. The two are complementary, not mutually exclusive.

Does MemGPT/Letta require cloud LLMs?

+
Letta supports both cloud LLMs (OpenAI, Anthropic, etc.) and local LLMs (via Ollama). SuperLocalMemory Mode A requires no LLM at all — retrieval is purely mathematical. This is the fundamental architectural distinction: SLM's core intelligence is information geometry, not language model inference.

What is the LoCoMo benchmark comparison?

+
Letta scores approximately 83.2% on LoCoMo, requiring cloud LLM calls. SuperLocalMemory V3 scores 74.8% in Mode A (zero cloud) and 87.7% in Mode C. SLM Mode C slightly outperforms Letta while SLM Mode A achieves near-comparable scores with zero cloud dependency.

Which should I use for building a production AI agent?

+
They serve different purposes. Letta is an agent framework with built-in memory management — you build agents in Letta. SuperLocalMemory is a memory layer — you add it to any agent (including Letta agents) via MCP. If you need local-first, EU-compliant memory for any agent framework, SuperLocalMemory is the memory backend. If you want a complete agent framework with memory management built in, Letta is a full solution.

Explore the Research

SuperLocalMemory's mathematical techniques — Fisher-Rao, sheaf cohomology, Langevin dynamics — are open source and designed to be adopted by any memory architecture, including Letta.

Research Papers Full Landscape Documentation