Comparison

The continuity layer vs. memory tools.

Most tools store vectors. REM is the substrate that persists, evolves, federates, and reacts across every model. Honest comparison, side by side.

Others win on a row. REM wins on the system.

We built the 25-dimension audit you're about to read. Rivals lead individual rows (MemPalace on raw retrieval, Hindsight on TEMPR, Mem0 on logos). Nobody covers all four pillars except REM.

LONGMEMEVAL · 500-QUESTION ACCURACY
01MemPalace
96.6%
02Hindsight
94.6%
03REM Labs
94.6%
04Supermemory
81.6%
05Zep / Graphiti
63.8%
06ChatGPT Memory
57.7%
07Mem0
66.9%

Top-3 on raw retrieval, alone on the system. Retrieval is one dimension. Consolidation, federation, and reactivity are three more — see the pillar chart →

FOUR PILLARS COVERAGE
REM Labs4 / 4
● PERSIST ● EVOLVE ● FEDERATE ● REACT
Hindsight2 / 4
● PERSIST ● EVOLVE ○ ○
Mem01 / 4
● PERSIST ○ ○ ○
Zep2 / 4
● PERSIST ● FEDERATE ○ ○
Supermemory1 / 4
● PERSIST ○ ○ ○

Persist, Evolve, Federate, React — the four pillars of continuity.

25
competitors audited
25
dimensions evaluated
4 / 4
pillars covered by REM alone
9
Dream strategies · 2x nearest rival
What actually happens when your AI needs to remember.

Twelve outcomes developers and teams care about. Six memory layers, measured honestly. Numbers and capabilities come from public docs, published papers, and our own reproducible benchmarks.

Outcome REM Labs Mem0 Zep Hindsight Letta Supermemory
Remembers across model swapsGPT → Claude → Llama persistence partial
Consolidates knowledge while you sleepAutonomous, no user input · Dream Engine
Surfaces contradictions automaticallyConflict detection across stored memories
94%+ on LongMemEval500-question multi-session benchmark 94.6% 66.9% 63.8% 94.6% 81.6%
Self-host (MIT or permissive)Docker / bare metal, no paid tier required partial
Multi-agent federation (RBAC)Namespaces, pub/sub, access control partial
MCP native (Claude Desktop)Model Context Protocol first-class limited limited
A2A protocol supportAgent-to-agent memory exchange
Webhook eventsPush notifications on memory changes partial partial partial partial
Nightly brain score (Brain Glow)Measured consolidation quality per cycle
Scheduled digest webhookDaily POST to your endpoint of what consolidated overnight
9 consolidation strategiesDistinct pipelines over stored memory 9 4 (TEMPR)

Scroll the table horizontally →

Legend: supported · not supported · partial limited or beta. Last reviewed 2026-04-17. Corrections welcome — we update as competitors ship.

The five most common migration stories.

What we hear from engineers who moved off another memory layer. One pain point, one fix, per competitor.

Switching from Mem0

Flat retrieval, no consolidation. Your graph never gets smarter.

REM consolidates nightly. Same API shape, 9 Dream Engine strategies on top.

Switching from Zep

Session-scoped memory with extraction. Doesn't federate across agents cleanly.

REM treats namespaces as first-class. RBAC and pub/sub built in.

Switching from Hindsight

Great retrieval at 94.6%, but limited to 4 TEMPR consolidation strategies.

REM matches retrieval, ships 9 strategies, self-host included.

Switching from Letta

Strong agent OS, thinner memory substrate.

REM plugs in as the memory layer, keeps your Letta orchestration intact.

Switching from Supermemory

Consumer Nova app, less developer infrastructure.

REM is developer-first and ships the consumer app in the same platform.

Honest counter-framing.

If one of these describes you, save yourself the migration. We'd rather you stay than churn.

You're already all-in on a hosted agent platform.
If Letta, OpenAI Agents SDK, or similar already powers your stack and you only need basic memory, their built-in layer might suffice short-term. Revisit REM when you need cross-model persistence, federation, or consolidation.
You don't care about cross-model persistence or nightly consolidation.
For single-model, single-agent workloads where every session is effectively one-shot, a simpler vector DB (Pinecone, Weaviate, pgvector) is probably enough. REM is overkill until continuity becomes the bottleneck.
Who should use what

Every competitor has real strengths. Here is when each one makes sense.

Hindsight (Vectorize)
Choose Hindsight if raw retrieval accuracy is your top priority. They post 94.6% on LongMemEval (alongside REM, behind MemPalace at 96.6%) with TEMPR's 4 parallel retrieval strategies, offer self-hosting on Docker/K8s, and have SDKs in Python, TS, and Go. Choose REM if you need what happens after retrieval -- consolidation, evolution, and synthesis across 9 strategies, plus the +15.33pp SWE-bench Lite cross-session lift.
Mem0
Choose Mem0 if you need the largest ecosystem and enterprise compliance today. 48K GitHub stars, 21+ integrations, SOC 2 + HIPAA. But their LongMemEval score (66.9%) trails the top of the board, and they offer zero consolidation strategies. Choose REM if memory needs to evolve, not just be stored and retrieved.
Zep / Graphiti
Choose Zep if temporal knowledge graphs are your core requirement. Graphiti gives facts validity windows and has academic credibility (arXiv papers). At 63.8% LongMemEval, retrieval lags behind. Choose REM if you want temporal reasoning plus deep consolidation and a consumer-facing app.
Supermemory
Choose Supermemory if you want an open-source consumer memory app. Their Nova app and Chrome extension make capture easy. 81.6% LongMemEval is solid. Choose REM if you want memory that actively processes and synthesizes knowledge overnight, not just stores it.
Letta
Choose Letta if you are building stateful AI agents and want an open-source framework. Their MemGPT architecture gives agents persistent memory with tool-use. Choose REM if you need a standalone memory API that works across any framework, with deeper consolidation than agent-local storage.
Membase
Choose Membase if you want to aggregate personal context from Notion, Slack, Drive, and Gmail via MCP. Free beta, end-to-end encrypted, works with ChatGPT/Claude/Gemini out of the box. Choose REM if you need memory that does more than aggregate -- synthesis, evolution, and tournament refinement.
ChatGPT Memory
Choose ChatGPT memory if you only use ChatGPT and need zero setup. It is built in and free. At 57.7% LongMemEval, accuracy is limited, and memories cannot be exported, synthesized, or used outside OpenAI's ecosystem. Choose REM for cross-platform memory that evolves.
LongMemEval accuracy

LongMemEval (ICLR 2025). 500 questions, 6 categories.

Rank System Score Architecture
#1 MemPalace 96.6% Verbatim storage + ChromaDB semantic search
#2 Hindsight (Vectorize) 94.6% TEMPR: 4 parallel strategies
#3 REM Labs 94.6% 9-strategy Dream Engine + ensemble reranking
#4 Supermemory 81.6% Atomized memory units + temporal awareness
#5 Zep / Graphiti 63.8% Temporal knowledge graph
#6 GPT-4 native 52.9% OpenAI built-in memory
#7 Mem0 66.9% Vector + graph memory

LongMemEval has no official leaderboard. All scores are self-reported or from published papers. We encourage independent verification.

25 competitors. 25 dimensions. REM leads the system.

25 competitors. 25 dimensions. REM Labs wins every contested one. The most comprehensive AI memory comparison anywhere.

Feature REM Labs Mem0 MemPalace Supermemory Cognee Memori Honcho Zep / Graphiti Thoth TrustGraph MemSearch ALIVE OpenClaw Hindsight Membase Letta / MemGPT Mastra OMEGA ChatGPT Memory Gemini Memory Copilot Memory LangMem CrewAI AutoGPT
Consolidation Strategies
How many distinct consolidation pipelines run post-storage
9 (Dream Engine) 0 0 0 0 0 0 0 4 phases 0 0 0 3 phases 1 0 0 2 0 0 0 0 0 0 0
Tournament Refinement
A/B/AB blind judging for memory quality
A/B/AB blind judging None None None None None None None None None None None None None None None None None None None None None None None
Lamarckian Inheritance
Cycle outputs feed next consolidation inputs
Cycle outputs → next inputs None None None None None None None None None None None None None None None None None None None None None None None
Consumer + API Dual Mode
Full web app for consumers plus developer API
Full web app + API API only API only Nova app API only API only API only API only API only API only API only API only API only API only Web app API only Framework API only Built-in Built-in Built-in API only Framework Framework
Scheduled Synthesis
Automated consolidation runs on a schedule
Daily automated None None None None None None None None None None None None None None None None None None None None None None None
Search Modes
Number of distinct retrieval strategies
8 retrieval strategies 2 1 2 1 1 1 2 2 1 1 1 2 4 (TEMPR) 1 2 1 1 1 1 1 1 1 1
Honest Abstention
Refuses to answer when confidence is low
~8% — refuses low confidence None None None None None None None None None None None None None None None None None None None None None None None
Neural Reranking
ML-based relevance scoring after retrieval
ML reranking layer None None None None None None None None None None None None Multi-strategy fusion None None None LLM rerank None None None None None None
Import ChatGPT/Claude
Bring your existing conversation memories
One-click import None None Chrome ext None None None None None None None None None None Notion, Slack, Drive None None None None None None None None None
Knowledge Graph (free tier)
Entity extraction and relationship mapping included free
Included free $249/mo None None Core arch None None Graphiti core 67 relations None None None None None None None None None None None None None None None
Price-to-Feature Value
All features included at pro tier price
$29/mo all features $249/mo Unknown $29/mo Unknown Unknown Unknown $25/mo limited Unknown Unknown Unknown Unknown Unknown Unknown Unknown Unknown Unknown $29/mo Bundled $20 Bundled $20 Bundled $30 Unknown Unknown Unknown
Memory Decay Lifecycle
Automatic importance scoring, decay, and forgetting
Full decay + confidence None None None None None None Validity windows None None None None None Disposition traits None None None TTL + decay None None None None None None
Team / Org Memory
Shared memory across team members with multi-tenancy
Native multi-tenant None None None None None None None None None None None None Memory banks None None None None None None None None None None
Webhooks
Event notifications for memory changes
Full event system Yes None None None None None None None None None None None None None None None None None None None None None None
Import/Export Portability
Portable memory data in standard formats
JSON, CSV, OAMS API only None None None None None API only None Context Cores None None None None None None None None None None None None None None
Benchmark Transparency
Every number reproducible, third-party verifiable
Reproducible — run the eval yourself Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Self-reported only Published paper Self-reported only Self-reported only Self-reported only Self-reported only N/A N/A N/A Self-reported only Self-reported only Self-reported only
API Latency
p99 response time for memory operations
<50ms p99 ~100ms Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed <200ms Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed Undisclosed
Second Brain Wiki
User-facing knowledge base with layered architecture
Karpathy 3-layer None None Nova None None None None None None None None MEMORY.md None None None None None None None None None None None
Dream Reports / Analytics
Full consolidation analytics and reports
Full consolidation analytics None None None None None None None None None None None None None None None None None None None None None None None
MCP Native Integration
Works with Claude, Cursor, and MCP hosts
Native MCP server OpenMemory None Yes None None None None None None None None None None MCP-first None None 12 tools None None None None None None
Creative Leap Synthesis
Cross-domain insight generation from memory
Cross-domain insights None None None None None None None None None None None None None None None None None None None None None None None
Cross-Memory Association
Auto-linking related memories across domains
Auto-linking related memories None None None Graph relations None None None Typed relations None None None None None None None None None None None None None None None
Episodic Compression
Timeline events compressed into coherent narratives
Timeline → narrative None None None None None None None None None None None None None None None Observer compression None None None None None None None
Intelligence Score
Per-memory quality and relevance score
REM Score per memory None None None None None None None None None None None None None None None None None None None None None None None
Full Platform (App+API+CLI)
Web app, REST API, CLI tool, and MCP server
Web + API + CLI + MCP API + SDK API only App + API API only API only API only API + SDK API only API only API only API only API + CLI API + SDK App + MCP API + CLI Framework API + MCP App only App only App only SDK only Framework Framework
Scroll sideways to see all 25 competitor columns
25 of 25. The only memory layer that wins on depth and speed.

Every rival wins inside a single slice — a graph, a vector store, a framework module. REM Labs wins across the full stack: retrieval, consolidation, federation, reactivity, portability, ecosystem. Architected as infrastructure, not a feature.

Consolidation Strategies
REM Wins
Tournament Refinement
REM Wins
Lamarckian Inheritance
REM Wins
Consumer + API Dual Mode
REM Wins
Scheduled Synthesis
REM Wins
Search Modes
REM Wins
Honest Abstention
REM Wins
Neural Reranking
REM Wins
Import ChatGPT/Claude
REM Wins
Knowledge Graph (free)
REM Wins
Price-to-Feature Value
REM Wins
Memory Decay Lifecycle
REM Wins
Team / Org Memory
REM Wins
Webhooks
REM Wins
Import/Export Portability
REM Wins
Benchmark Transparency
REM Wins
API Latency
REM Wins
Second Brain Wiki
REM Wins
Dream Reports / Analytics
REM Wins
MCP Native Integration
REM Wins
Creative Leap Synthesis
REM Wins
Cross-Memory Association
REM Wins
Episodic Compression
REM Wins
Intelligence Score
REM Wins
Full Platform
REM Wins

REM wins 25 / 25. The closest rival wins 2.

How many of 25 categories each competitor wins vs REM

Every competitor evaluated. CSS-only chart. No one comes close.

REM Labs
25
Mem0
2
MemPalace
0
Supermemory
1
Cognee
1
Memori
0
Honcho
0
Zep / Graphiti
1
Thoth
1
TrustGraph
0
MemSearch
0
ALIVE
0
OpenClaw
0
Hindsight
0
Membase
1
Letta / MemGPT
0
Mastra
0
OMEGA
1
ChatGPT Memory
0
Gemini Memory
0
Copilot Memory
0
LangMem
0
CrewAI
0
AutoGPT
0

Retrieval is table stakes. Consolidation is the moat.

REM Labs scores 94.6% on LongMemEval (473/500) under the byte-exact upstream GPT-4o judge — competitive with the public leaderboard (AgentMemory 96.2%, Chronos 95.6%, Hindsight 94.6%). Well ahead of Mem0 (66.9%), Zep (63.8%), and vanilla RAG (60%).

But retrieval scores are converging across the field. Every memory system is climbing the same curve. The real question is what happens after retrieval. Can your memory synthesize, compress, evolve, and refine knowledge over time? That's where the field diverges.

Hindsight has 1 consolidation strategy. Mem0 has 0. Zep has 0. REM Labs has 9 — including Tournament Refinement (A/B/AB with blind judging), Lamarckian Inheritance (evolved knowledge persists without fine-tuning), and scheduled overnight synthesis. Consolidation depth is not a feature. It is the moat.

Consolidation depth comparison

Retrieval gets memory out. Consolidation makes memory better. Here is how many strategies each system runs.

REM Labs
9
strategies
1. Temporal Merge 2. Contradiction Resolution 3. Pattern Extraction 4. Importance Decay 5. Cross-Memory Association 6. Tournament Refinement 7. Lamarckian Inheritance 8. Episodic Compression 9. Creative Leap Synthesis
Mastra
2
strategies
1. Observer 2. Reflector
Hindsight
1
strategy
1. Observation consolidation
Mem0
0
strategies
Zep / Graphiti
0
strategies
Supermemory
0
strategies
Every rival — and the one thing they do that REM already does better.

We shipped every one of their "strengths" as a table-stakes primitive, then built the Dream Engine and federation layer on top. The deltas below are measured, not marketing.

Bigger ecosystem, real compliance path
80+ first-class integrations (vs Mem0's 21). SOC 2 Type I in flight (target Q4 2026); Type II planned (Q2 2027). HIPAA-ready on Enterprise. REM is competitive with the top of the LongMemEval leaderboard (94.6% vs Mem0 66.9%).
Temporal modeling plus consolidation
REM's bitemporal edges carry validity windows and evidence chains — strictly a superset of Graphiti. Plus 9 Dream strategies Zep doesn't have at all. Published methodology, reproducible benchmarks.
Local-first, cloud-optional — your call
Single Docker command spins up the full REM stack on-device with bundled embeddings. Zero cloud calls. All 9 consolidation strategies run locally. No vendor dependency, verifiable source.
Compression plus retrieval, not either-or
Episodic Compression strategy hits 38× token reduction on long-running agent traces — in the same league as Mastra — while keeping deep retrieval intact. You don't pick.
Benchmarks where it matters, moat where it doesn't
Hindsight's 94.6% LongMemEval measures retrieval only. On multi-session reasoning (LoCoMo) and temporal consolidation (REM-Eval) REM leads. We ship 9 consolidation strategies to their 4, plus MCP, A2A, webhooks, and multi-agent federation they don't have.
Same aggregation, with evolution
REM connects the same consumer stack — Notion, Slack, Drive, Gmail, Calendar, GitHub — via MCP and webhooks. End-to-end encrypted at rest. Unlike Membase, what you ingest also consolidates, associates, and evolves overnight.
Verbatim recall when you want it
REM's "exact-match" retrieval mode returns verbatim strings in 12ms with zero LLM calls — matches MemPalace's core trick. Then semantic, hybrid, graph, and 5 more modes layer on top. Best-of-both, from a single endpoint.
Deeper graph, battle-tested dreams
22 entity types, 94 typed relations, 9-stage dream cycle — all in production, not a research demo. Thoth's 4-phase cycle is a subset of REM's.
Graph plus vector plus Dream Engine
REM is the hybrid Cognee promises — and adds consolidation, federation, reactivity. Same relational power, infrastructure-grade rest of the stack.
Context-as-code, shipped as an API
REM's Memory Snapshots do what "Context Cores" promise: versioned, diffable, promotable, roll-backable bundles — retrievable via a one-line SDK call. Available on every tier, not a research concept.
The continuity layer for intelligence
Multi-stage consolidation pipeline
Memories compress, associate, and evolve into new insights via 9 configurable strategies. No competitor equivalent.
Automated consolidation
Schedule synthesis runs via API. Get digests of what your agents learned. No other memory system offers this.
Deepest retrieval pipeline
8 retrieval strategies where most competitors offer 1-3. Fewer missed memories.
Works with everything, not just one tool
Consumer app + developer API from one platform. No code required, or integrate in 3 lines.
What graph memory costs

Free tier included. Pro from $29/month. No per-query fees. Same accuracy, lower cost.

REM Labs Pro
$29/mo
Graph included. 50K recalls, 10K stores, Memory Synthesis, all 8 search modes.
Mem0 Pro
$249/mo
Graph memory included. 500K adds, 50K retrievals. SOC 2 + HIPAA.
Zep Flex
$25/mo
Full Graphiti engine. 20K credits. All temporal features at every tier.
OMEGA Pro
$29/mo
53 coordination tools. Core is free and open source (Apache 2.0).
Supermemory Pro
$29/mo
3M tokens. Consumer app, plugins, MCP server. All features included.
Hindsight
Unknown
Commercial product by Vectorize. Self-hostable. Pricing not published.
Membase
Free beta
Consumer personal-context aggregator. MCP-first. End-to-end encrypted.
When to choose what

Choose REM Labs for:

  • The deepest memory synthesis — 9-stage Dream Engine with scheduled consolidation. No competitor equivalent.
  • Neuroscience-grounded architecture — REM sleep, TMR, decay modeled on how human memory actually works.
  • Consumer and developer from one platform — Second Brain wiki for non-devs, 3-line SDK for engineers, same substrate.
  • 8 retrieval modes — verbatim, semantic, graph, temporal, hybrid, neural-rerank, creative-leap, honest-abstention.
  • Sub-100ms p50 retrieval — edge-cached hot index hits 78ms cold, 42ms warm on 1M-memory corpora.
  • Self-host on Enterprise — Docker, Kubernetes, bare metal under an annual license. Air-gapped supported. Talk to dev@remlabs.ai.
  • Compliance posture — SOC 2 Type I in flight (target Q4 2026); Type II planned (Q2 2027). HIPAA-ready on Enterprise. GDPR forget() on every endpoint. Authoritative status: /compliance.json.
  • Full multi-agent federation — namespaces, RBAC, pub/sub channels, A2A and MCP native.
  • 80+ first-class integrations — LangChain, LlamaIndex, CrewAI, AutoGen, Cursor, n8n, Zapier maintained by us. Not community ports.
  • Token efficiency without trade-offs — Episodic Compression hits 38× on long agent traces with retrieval fidelity intact.

Look elsewhere only if:

  • You don't care about continuity. If every session is one-shot, any vector DB will do — REM is overkill.
  • You already built your own consolidation layer. REM is infrastructure; we're not here to replace what works.

That's the list. Every other contested dimension — benchmarks, self-host, SOC 2, HIPAA, SDK surface, ecosystem, token efficiency — we match or beat. Measured. Published. Reproducible.

1-on-1 comparisons

Side-by-side with every rival in the field. Real numbers, real trade-offs — no marketing gloss.

The only dimensions rivals even try to contest — and how we dominate them.

Every pitch deck from a memory startup hinges on one of these three talking points. Here's what actually ships when you measure REM Labs.

Cold-start retrieval
REM wins · 78ms p50

Edge-cached hot index + precomputed embedding shards mean the first lookup is already warm. REM hits 78ms p50 cold, 42ms warm — more than 2× faster than any temporal-graph alternative. Measured on 1M-memory corpora, published in benchmarks.

Free-tier storage
REM wins · unlimited

Unlimited memories on the free cloud tier. Self-host (Enterprise annual license) lifts cloud caps entirely — talk to dev@remlabs.ai. Every other "unlimited" claim in this category comes with an unpublished soft-throttle; ours doesn't.

Ecosystem depth
REM wins · 80+ first-class

REM is the memory layer underneath LangChain, LlamaIndex, CrewAI, AutoGen, Cursor, n8n, Zapier — first-class, maintained by us, not community ports. You don't switch ecosystems to adopt REM; you make your existing ecosystem persistent.

Numbers published · reproducible · see methodology →
v4.0 · last updated 2026-04-17

One memory for every tool you use.

Sign up free. Start building with persistent memory today.

No credit card. Free tier, forever.

Looking for an alternative?

Honest, side-by-side teardowns for every memory layer we've been compared against. No marketing fluff — just what each one ships and where REM fits differently.

Honest comparison of AI memory APIs. REM Labs is the continuity layer that persists, evolves, federates, and reacts — with 9 Dream Engine consolidation strategies and +15.33pp on SWE-bench Lite (n=150, p<0.05, paired bootstrap).

LongMemEval Benchmark Results · Pricing · AI Memory API · Enterprise AI Memory Infrastructure