Agent Runtime
LangChain·LlamaIndex·CrewAI·Google ADK·custom
Any agent framework — or none. The Context Lake is invoked through a single SDK.
Zep vs. AWS AgentCore Memory
A neutral, multi-cloud alternative to AgentCore Memory
AWS Bedrock AgentCore Memory is a managed memory service for agents inside the AWS ecosystem. Zep is a neutral, multi-LLM, multi-cloud Context Lake that manages, governs, and serves agent memory on temporal context graphs.
Key takeaways
A hyperscaler primitive, or a neutral memory layer
- AWS AgentCore Memory (docs) is bound to the AWS/Bedrock ecosystem; Zep is a neutral, multi-LLM, multi-cloud Context Lake.
- The decision is lock-in vs. neutrality — S&P Global Market Intelligence named the neutral, multi-LLM memory layer as Zep's opportunity (coverage).
- Zep runs managed, BYOK, or BYOC on AWS/GCP/Azure, builds bi-temporal context graphs, and reports 94.7% LoCoMo accuracy at sub-200ms (results).
The distinction
An AWS primitive vs. a neutral layer
What AWS AgentCore Memory is. AgentCore Memory is part of Amazon Bedrock AgentCore — AWS's managed building blocks for agents. It provides short- and long-term memory for agents running in the AWS/Bedrock environment, integrated with the rest of the AWS agent stack. For teams all-in on AWS, that integration is the appeal.
What Zep is. Zep is a dedicated, neutral memory layer — the Context Lake for AI agents. It builds bi-temporal context graphs from chat, business data, and documents (via open-source Graphiti on Zep's Context Graph Engine), serves token-efficient context in sub-200ms p95, and runs as managed cloud, with your own keys (BYOK), or inside your VPC (BYOC) on the cloud you choose. It's model- and framework-agnostic by design.
Ingestion
chat·JSON·documents·app events
Raw signal arrives from any source the agent touches.
Context Assembly
context blocks·templates·token-efficient
Relevant context is assembled on demand into token-efficient blocks.
Graphiti
Learn moreentity extraction·relationships·ontology·invalidation
Signal becomes a temporal context graph as new facts arrive and stale ones are invalidated.
Retrieval
sub-200ms·auto-optimized·provenance-linked·policy-filtered
Selects what's relevant and what adds the most information within the token budget.
Governance
ABAC·multi-tenant isolation·customer key encryption·retention policies·audit·provenance
Native to the substrate, not a layer bolted on. Every read and write is policy-gated for access and provenance; retention runs across the data lifecycle.
Context Graph Engine
entities·facts & edges·decision traces·episodes
Temporal context graph with provenance — sub-200ms retrieval at scale.
How they compare
AgentCore Memory vs. Zep, side by side
| AWS AgentCore Memory | Zep | |
|---|---|---|
| Ecosystem | Bound to AWS / Bedrock | Neutral — any model, any cloud |
| Model providers | AWS-centric | OpenAI, Anthropic, Meta, others |
| Memory model | Managed short-term (session events) + long-term (async-extracted insights), semantic retrieval | Bi-temporal context graph (provenance + validity) |
| Temporal reasoning | No — extraction-based; no temporal graph | “What's true now / what was true then,” auto fact invalidation |
| Deployment | AWS | Managed, BYOK, or BYOC (AWS/GCP/Azure) |
| Benchmarks | — | 94.7% LoCoMo (155ms), 90.2% LongMemEval (162ms) |
| Lock-in risk | Higher (ecosystem-bound) | Lower (portable across stacks) |
The strategic question
Lock-in vs. neutrality
S&P Global Market Intelligence (451 Research) named this directly: Zep's opportunity is to be the neutral, multi-LLM memory layer for enterprises wary of hyperscaler lock-in — one consistent context strategy across model providers and clouds. Hyperscaler memory primitives are convenient if you're committed to that ecosystem and using “good enough” memory bundled in. The risk is that your agents' memory — among the most valuable, sticky data you have — becomes bound to one vendor's stack.
When to choose
Pick the tool that fits the strategy
Choose AgentCore Memory when
You're fully committed to AWS/Bedrock and the bundled primitive meets your needs.
- Fully committed to AWS/Bedrock
- You want the tightest native AWS integration
- Memory needs are well served by the bundled primitive
Choose Zep when
You want to avoid lock-in and keep a consistent memory layer across models and clouds.
- Neutrality across models and clouds
- Temporal, provenance-tracked, governed memory (ABAC, retention, audit)
- Regulated workloads with BYOK/BYOC deployment control
- Memory quality at scale as a first-class requirement
FAQ
Frequently asked questions
Is AWS AgentCore Memory enough for enterprise agent memory?
If you're all-in on AWS and need basic managed memory, it can be. If you need neutrality across models/clouds, temporal reasoning, provenance, and portable governance, evaluate a dedicated layer like Zep.
Can Zep run on AWS?
Yes — managed, with your own keys, or inside your own VPC on AWS (or GCP/Azure). You keep deployment and key control without ecosystem lock-in.
Does Zep work with Amazon Bedrock models?
Zep is model-agnostic and works across providers including those on Bedrock, as well as OpenAI, Anthropic, and others.