Side-by-Side · Updated 2026-04-17

REM Labs vs MemGPT
what's different, what's honest, when to pick which.

MemGPT (now Letta) pioneered hierarchical memory tiers — main context + archival storage, with the LLM acting as its own OS-style memory manager. REM Labs picks up where that research stopped: nine consolidation strategies, production infrastructure, federation, reactivity. Honest side-by-side.

MEMGPT: RESEARCH PROJECT · HIERARCHICAL TIERS · SELF-HOST FOCUS · BSD LICENSE

What MemGPT does well

Credit where it's due.

MemGPT is a Berkeley research project that put "LLM as operating system" on the map. Without this paper, half the current memory ecosystem doesn't exist.

Foundational hierarchical memory

Main context / archival storage / recall buffer — the first published design where the LLM manages its own memory tiers via function calls. Elegant, well-documented, cited everywhere.

Self-host-first ethos

BSD license, pip-installable, runs fully local with Ollama or a hosted OpenAI endpoint. No vendor lock-in. Academic-grade reproducibility of the original paper.

Agent OS primitives, openly

Tools, sleep-time agents, core memory blocks — all explicit primitives you can reason about. If you're building agents from first principles, MemGPT is one of the cleanest reference implementations.

What REM does differently

Research tier → production continuity layer.

MemGPT gives you two memory tiers and a buffer. REM gives you nine consolidation strategies that run overnight, federation across agents, and webhook reactivity — all as a managed service or self-hosted, your choice.

SYNTHESIZEMerge related memories into higher-order insights.

PATTERN EXTRACTDetect recurring themes and behavioral signatures.

CONTRADICTIONFlag conflicting facts before they poison retrieval.

COMPRESSSummarize stale long-form content without losing semantics.

ASSOCIATEBuild implicit graph edges between memories.

VALIDATECheck facts against prior evidence and sources.

EVOLVERewrite summaries as new context arrives.

FORECASTPredict next-need memories before the user asks.

REFLECTSelf-audit retrieval quality and tune weights.

Production-grade vs research-grade

MemGPT's hierarchical tiers are brilliant in theory and most teams re-implement them when they move to production. REM ships all nine strategies as a managed service (or a Docker container), plus webhooks, MCP, A2A, GDPR endpoints, and multi-agent namespaces. Infrastructure primitives, not just reference code.

Side by side

Fourteen dimensions. Sourced, dated, honest.

MemGPT is primarily a self-host OSS project. We compared against the current release (Letta fork). Numbers on retrieval are from their own evaluation harness where available.

Dimension	REM Labs	MemGPT / Letta
Category	Continuity layer for intelligence	Hierarchical agent memory framework
LongMemEval (500q)	94.6% · byte-exact upstream GPT-4o judge	Not publicly reported on LongMemEval; prior self-evals on LIME / DMR
Consolidation strategies	9 (Dream Engine)	2 tiers + recall buffer (main / archival, self-managed)
Model-agnostic	Yes — every vendor + local	Yes — OpenAI, Anthropic, local (Ollama, vLLM)
Self-hostable	Yes — single Docker command	Yes — pip install, native strength
Open source	Partial — SDK + self-host OSS, Dream Engine hosted	Yes — BSD (full)
GDPR / forget API	Yes — per-memory + per-namespace + audit	Partial — delete from archival; no audit log
Federation across agents	Yes — shared namespaces + A2A agent card	Partial — shared blocks across agents (Letta)
Webhooks / reactivity	Yes — memory.created, dream.completed, contradiction.detected	No native; sleep-time agents can poll
MCP / A2A protocol	Yes — /.well-known/mcp.json	Partial — Letta adding MCP; A2A in roadmap
Multi-agent / hive	Yes — DreamHive, coordinated consolidation	Partial — multi-agent in Letta, no coordinated consolidation
Managed cloud	Yes — free tier, $19 Pro	Partial — Letta Cloud (closed beta)
Pricing start	Free (unlimited memories, 500 dreams/mo)	Free (self-host) · Letta Cloud: $0 free, paid TBD
Funding / project status	Solo founder, 2026	Berkeley research → Letta (seed round, team of ~10)

LONGMEMEVAL METHODOLOGY · /benchmarks

Head-to-head

vs. MemGPT / Letta.

The two dimensions MemGPT markets hardest — and REM's actual numbers on each.

Open source, end-to-end

REM's core is Apache 2.0 — SDKs, extractors, self-host edition with the full Dream Engine, 9 strategies, 8 retrieval modes. Self-host in one command, ~90s, unlimited everything. On-prem or air-gapped, no closed dependency.

Apache 2.0 core, Docker + K8s + bare metal.
Dream Engine runs on your hardware with your keys.

Inspectable agent primitives

REM exposes typed primitives: memory namespaces, RBAC, pub/sub channels, 9 consolidation strategies, 8 retrieval modes, tournament refinement, Lamarckian inheritance, webhook events. Every layer inspectable; every layer pluggable.

First-class MCP, A2A agent card, typed webhooks.
Researcher-grade hooks; production-grade defaults.

How to pick

Simple decision tree.

Pick REM if…

You want consolidation out of the box — not just tiers, but synthesis, contradiction-detection, forecasting.
You need a managed service with SLAs, not just a self-host package.
You need webhooks and reactivity baked into the memory system.
You want protocol-native shipping (MCP, A2A, OpenAPI, channels) day one.
You're shipping to production, not writing a paper.

Pick MemGPT / Letta if…

You specifically want the academic reference implementation of hierarchical main/archival tiering and nothing else.
You're writing a paper, not shipping a product.

They compose

You can run a MemGPT agent with REM as its consolidation layer. MemGPT handles the core/archival tiering; REM runs the nine Dream strategies on its archival corpus overnight. Both published, both documented. Email us if you want the integration recipe.

Continuity, production-grade, today.

No credit card. Python & Node SDKs. Self-host in 90 seconds or try the hosted free tier.

Try REM free View 25-competitor comparison →

Honest comparison policy · email hey@remlabs.ai if any row is wrong — we fix in 48h.