Source-linked capability map. Letta's product boundary is taken from its primary documentation. Protocol-scoped LoCoMo evidence is disclosed. No pricing comparisons.
SuperLocalMemory and Letta solve different problems. SLM is a standalone memory layer. Letta is a stateful agent runtime with memory hierarchy built in. They are largely complementary — verify each product boundary independently.
SLM exposes the same local memory system through four surfaces — MCP, CLI, hooks, and dashboard — without requiring a change to the agent runtime. Letta is the runtime: the LLM manages memory operations from inside the agent loop.
A memory layer with MCP and CLI surfaces, a SQLite-backed local core, nine framework adapters, bounded loops, RBAC, and GDPR controls. Plugs into any agent via MCP. Framework compatibility requires a tested adapter.
A complete stateful agent runtime. The LLM agent itself manages memory — deciding what to store, retrieve, and summarize — through in-context blocks, recall storage, and archival memory. Includes execution environment, not just a memory backend.
Product boundary taken from Letta's primary documentation. MemGPT was rebranded as Letta in 2024. Verify current behavior from the versioned release documentation before making deployment decisions.
In-context memory blocks, archival memory, and recall storage. The LLM agent manages memory operations — store, retrieve, summarize, compress — through function calls within the runtime. Supports local LLMs (Ollama) and cloud providers. Letta Cloud provides a managed deployment.
Choose Letta when: Building full agents inside a dedicated runtime with built-in LLM-managed memory, or when Letta-native tooling, multi-agent orchestration, and the Letta server deployment model fit your use case.
Read primary source →Choose SLM when your agents need a standalone memory layer that plugs into any framework via MCP: SQLite-backed persistent memory, three explicit operating modes, nine npm framework adapters, bounded loops with independent-gate verification, per-workspace RBAC, GDPR Art. 15/17/20 controls, and four integration surfaces — without requiring a change to the agent runtime.
Mode A keeps the core memory path local after required assets are present. Mode B adds a configured Ollama endpoint. Mode C adds a configured cloud provider. Optional connectors, proxies, backups, and clients have separate network paths that require independent assessment.
Attributes are configuration-dependent. Review each item independently for your deployment. Letta values are sourced from primary documentation; verify current behavior from the versioned release.
| Capability | SLM V4 | Letta (formerly MemGPT) |
|---|---|---|
| What it is | Standalone memory layer via MCP / CLI / hooks / dashboard | Stateful agent runtime with memory hierarchy built in |
| Memory management | Five candidate producers, fusion, optional reranking, graph + Fisher-informed score enhancement | LLM-managed: agent decides store / retrieve / summarize via function calls |
| Storage layer | SQLite-backed, local-first; configurable data root | PostgreSQL-backed Letta server; verify selected deployment |
| Operating modes | A (local sentence-transformer) / B (Ollama) / C (configured cloud) | Provider and deployment-dependent; verify from release documentation |
| Framework adapters | 9 npm packages (V4) | Letta Python SDK / REST API / Letta-native agents |
| Teams & RBAC | Admin / member / viewer; per-workspace isolation | Letta Cloud organization-level; verify self-hosted configuration |
| Bounded loops | Independent-gate verified (V4) | Agent orchestration via Letta runtime |
| GDPR controls | Art. 15/17/20; hash-chained audit trail; opt-in PII redaction | Verify provider and deployment terms |
| EU AI Act self-assessment | Per-mode (self-assessment, not certification) | Not documented |
| Published preprints | 3 arXiv preprints (2603.14588, 2603.02240, 2604.04514) | MemGPT research paper (2310.08560) |
| License | AGPL v3 | Apache 2.0 |
Documented in the V4 release. Absent from or not documented in most agent memory systems — verify each for your deployment.
The canonical memory source is SQLite-backed at a configurable data root. No cloud persistence is required for the core memory path in Mode A or B.
Three documented operating modes with distinct embedding paths and network behaviors. Mode can be switched at runtime — memories persist across the transition.
Nine adapter packages published on npm in V4. Adapters provide framework-specific bindings without requiring a rewrite of existing tool configuration.
V4 adds bounded loop orchestration with an independent verification gate at each iteration boundary. Gate results are inspectable via CLI and dashboard.
Admin / member / viewer role-based access with per-workspace isolation. Single-user setups are unaffected — no login required for solo use.
Personal, shared, and global memory boundaries within the same runtime. Namespaces can have independent mode configurations for different projects.
GDPR Art. 15/17/20 controls (access, erasure, portability), a hash-chained audit trail, and opt-in PII redaction. A per-mode EU AI Act self-assessment ships with the tool — a technical-control map, not a legal certification.
Four integration surfaces — Model Context Protocol, structured CLI, event hooks, and the multi-agent dashboard — without reconfiguring existing workflows.
Four public preprints with protocol-scoped LoCoMo evidence and architecture documentation. See Research →
Published V3 results carried into V4 with original protocol scope. Not a fresh V4 rerun. Letta does not publish LoCoMo results on matching protocol — direct ordinal comparison is not possible.
10 conversations / 1,276 questions. Local embeddings, local retrieval. No LLM answer construction.
10 conversations / 1,276 questions. Local retrieval with GPT-4.1-mini answer synthesis disclosed.
Conv-30 only / 81 questions. text-embedding-3-large plus GPT-4.1-mini generation and judge.
Published V3 architecture evidence carried into V4. Mode A covers 10 conversations and 1,276 questions; the 74.8% retrieval result discloses GPT-4.1-mini answer synthesis. Mode C covers Conv-30 only (81 questions) with cloud embeddings and GPT-4.1-mini. Scope differences between Mode A and Mode C make direct ordinal comparison unreliable. LoCoMo protocols are not comparable without matching dataset, answerer, judge, prompts, and context budget.
SuperLocalMemory ships a per-mode EU AI Act self-assessment. It is a technical-control map, not a legal certification — applicability depends on your deployment, data, and operator role.
Regardless of mode: GDPR access / erasure / portability (Art. 15, 17, 20), a hash-chained audit trail, per-workspace isolation, opt-in PII redaction, and admin / member / viewer role-based access. See Governance & EU AI Act controls →
Open source, AGPL v3. A Qualixar Research Initiative.