Sentra vs Zep: Company Brain vs Agent Memory Graph

TL;DR

• Zep is an enterprise agent memory layer built on Graphiti, organizing context into per-entity graphs (user, agent, org, domain) with sub-200ms retrieval.
• Sentra is a company brain, one org-wide graph that humans and every agent read and write, built through write-time comprehension against a per-org ontology.
• Zep isolates memory by entity across 1.4M+ graphs. Sentra unifies it, so what you teach one agent, every agent and teammate remembers.
• Zep serves memory to agents. Sentra serves a shared brain to the whole company and its agents.
• Choose Zep for agent-centric, high-throughput workloads where isolation and speed dominate. Choose Sentra when stale or contradictory knowledge across humans and agents is the liability.

What Each Tool Actually Does

Memory architecture decides what your AI can know and who gets to know it. A retrieval layer built for one agent serves that agent well and tells the rest of the company nothing. A shared graph spreads context across every person and every model but demands a common structure to do it. Zep and Sentra both store facts and both track time, yet they answer different questions about scope and audience.

Zep is an agent memory layer built on Graphiti, a temporal knowledge graph engine. It organizes context into per-entity graphs by user, agent, org, and domain, and its dashboard example shows over 1.4 million active graphs running as one Context Lake. Retrieval stays under 200ms at p95 even at 100 million nodes, which makes it well suited to high-throughput agent workloads where each entity's memory stays isolated by design.

Sentra is a company brain. Every team member and every agent reads and writes to one org-wide graph, and the promise is direct. What you teach one agent, every agent remembers. Sentra resolves meaning at ingestion through write-time comprehension, building the graph against an ontology unique to your organization rather than guessing from vector similarity at query time. The audience is the whole company and its agents, not a single agent in isolation.

How This Comparison Was Made

We scored both systems on seven criteria: graph scope, audience fit, temporal modeling, retrieval performance, commitment and contradiction handling, governance, and deployment. Scope and audience fit carry the most weight, because they shape what every downstream feature can and cannot do. The rest break ties between two systems that both model time and both run at enterprise scale.

Two gaps limit any analysis here. Neither Zep nor Sentra publishes pricing tiers, so this comparison makes no cost claims beyond what each vendor states. Sentra's MEME benchmark scores come from KAIST's 2026 evaluation. Where a number comes from one vendor, we attribute it. Everything else rests on each company's documented architecture.

Side-by-Side: Zep vs Sentra at a Glance

The table below maps both systems across the dimensions that decide which one fits your stack.

Graph Scope: Per-Entity Graphs vs One Org-Wide Graph

Graph scope is the decision that determines everything else. Zep and Sentra both build temporal knowledge graphs, but Zep partitions memory into millions of isolated graphs while Sentra runs one shared graph for the whole organization. That single choice shapes who can read each fact, how isolation works, and what the system can notice.

Zep organizes memory into four entity types, each kept as its own graph. A user graph, an org graph, an agent graph, and a domain graph track entities and facts separately, and the dashboard example shows 1,402,891 active graphs across the Context Lake. Isolation is the point. One agent's memory stays walled off from another's, which keeps tenant data clean and lets the system scale horizontally as you add entities. For agent-centric workloads where each agent serves a distinct user or task, that boundary is a feature, not a limit.

Sentra makes the opposite bet. Every team member and every agent reads and writes the same graph, so what you teach one agent, every agent remembers. A fact a sales agent learns about a customer is immediately available to a support agent, an analyst, and a human reading the same brain. Cross-entity questions become answerable because the facts live in one place rather than scattered across separate stores.

The shared model demands something Zep's per-entity design avoids. A single graph only stays coherent against a unified ontology, so Sentra resolves semantics at write time against an ontology unique to your organization. You pay for cross-agent recall with the work of defining what your entities mean. Zep skips that cost and gives up cross-entity recall in return. Choose the scope that matches whether your agents need to share knowledge or stay isolated.

Temporal Modeling: Fact Invalidation vs Bi-Temporal Awareness

Both systems track time, and both keep history rather than overwriting facts. The difference lies in how many timestamps each fact carries, and what that buys you when an agent reads it back.

Zep attaches a validity window to each fact. When new information contradicts an existing fact, Zep invalidates the old one and retains it as history rather than deleting it. An agent can query what is true now or what was true on any past date, so a fact like "Emily prefers cycling to jogging (Valid: 2024-11-14 — present)" stays queryable even after Emily's preference shifts (getzep.com). The model tracks when a fact was recorded and when its validity ended, which is enough for most agent context that asks "what is the current state for this entity?"

Sentra stamps every fact twice. One timestamp marks when the fact became true, and the other marks when it stopped being true. Old facts are invalidated, never deleted, and provenance is treated as first-class. That second timestamp is the wedge. A bi-temporal graph knows both when a fact entered the system and the real-world window during which it held, so an agent can reconstruct what the company believed at a point in time versus what was actually true then.

The practical payoff shows up against flat embedding stores. In a vector index, old facts sit next to new ones, equally weighted, ready to restate yesterday as today. Sentra resolves this by weighting facts against their temporal validity, so a deprecated pricing tier or a reversed decision does not surface as current. If your agents act on commitments, policies, or numbers that change and get reversed, the second timestamp prevents them from quoting stale truth as live.

Retrieval and Ingestion

Zep optimizes for speed at scale, and the numbers back it. Retrieval stays sub-200ms p95 whether a graph holds 10,000 nodes or 100 million, with p95 latency moving from 148ms to 168ms across that range (getzep.com). On the LoCoMo benchmark Zep reports 94.7% accuracy at 155ms retrieval, and on LongMemEval 90.2% at 162ms (getzep.com). Ingestion is just as tight. A single add_messages call with return_context=True writes new memory and returns context in one round trip, which keeps agent loops fast and simple to build.

Sentra makes a different trade at write time. Rather than store text and resolve meaning when an agent queries, Sentra resolves semantics at ingestion against an ontology unique to your organization. Vector search returns what is close, not what is correct, so Sentra treats meaning as a primitive rather than a query-time guess. That comprehension happens once on the way in, which is why a fact arrives in the graph already typed, dated, and connected to its source.

The trade-off is real on both sides. Zep's single-call API and flat latency curve reward agent-centric workloads where you want answers in milliseconds and isolation per entity. Sentra's write-time work costs more at ingestion but pays off when many agents and humans read the same fact and need it to mean the same thing every time. Sentra exposes the graph over REST or MCP and connects to 200+ tools including Slack, Gmail, HubSpot, and GitHub. On the MEME benchmark, Sentra is the only system above 30% on both Cascade and Absence.

Commitment Tracking and Contradiction Detection

Sentra treats commitments as objects with a lifecycle, not as facts buried in chat history. Its action memory tracks a promise from the moment someone says it, keeps the evidence attached, and reports follow-through status. One operations example shows six tracked commitments across four partners: four shipped, one slipped on SAML SSO, and one quietly dropped. Contradiction detection runs on top of the same graph and surfaces reversals and stale facts. An executive view cites nine material decisions, three reversals, and two unfulfilled commitments since the last board meeting, with the reversals concentrated in pricing.

Zep does infer patterns, and its Observations feature is genuine signal. It analyzes graph structure to surface recurrences like "Jane has upgraded within two weeks of each of the last three product launches," derived from three dated upgrade events (getzep.com). That works well inside a single entity graph where the data lives together.

The reason Zep does not surface cross-entity drift is architectural, not a missing feature. Its memory splits into separate graphs per user, agent, org, and domain, and a commitment made in a Slack thread, a contradiction between a sales rep and an engineer, or a reversal that spans two teams lives across graphs that Zep keeps isolated by design. Isolation gives Zep its retrieval speed and clean per-agent context, but no single graph holds the cross-party view that commitment tracking needs. Sentra's one org-wide graph holds every promise and every fact in the same place, so a contradiction between two people becomes detectable rather than scattered.

Governance, Security, and Deployment

Both systems clear the enterprise security bar, but they diverge on isolation depth and compliance scope. Zep runs SOC 2 Type II and offers a HIPAA Business Associate Agreement, which matters if you handle protected health data. Its access model is attribute-based, so you grant or deny reads against fact-level attributes rather than coarse roles. Retention is policy-driven and supports legal hold, and every fact traces back to its source episode with a timestamp.

Zep's deployment range favors regulated cloud workloads. You can bring your own key on AWS, GCP, or Azure, or run Zep entirely inside your own VPC with bring-your-own-cloud. Both options keep encryption keys or the full deployment under your control.

Sentra carries SOC 2 Type II and adds ISO 27001, which buyers in Europe and procurement-heavy enterprises often require. It deploys to cloud, an isolated VPC, or a fully air-gapped on-prem environment with no outbound connectivity. The air-gapped option matters when a defense contractor or financial institution cannot let company knowledge touch a vendor network.

The split is practical. Zep covers HIPAA workloads with attribute-based controls. Sentra covers ISO 27001 mandates and true air-gapped isolation.

Best For: Zep

Choose Zep when you build agent products and your hardest constraint is fast, isolated retrieval at scale. Zep keeps p95 latency under 200ms even at 100 million nodes, and it organizes memory into separate USER, AGENT, ORG, and DOMAIN graphs so each agent or customer holds its own context without bleeding into another. For a multi-tenant assistant serving thousands of accounts, that isolation is the point, not a limitation.

The decision signal is throughput per entity. If you run high-volume agent workloads where each user or tenant needs a clean, queryable history and millisecond retrieval matters more than shared organizational context, Zep fits. Its single-call add_messages with return_context=True writes and reads in one round trip, which keeps agent loops tight. Zep reports 94.7% on LoCoMo and 90.2% on LongMemEval, with governance built at the substrate level through attribute-based access control and policy-driven retention. For developers shipping agents, that combination is hard to beat.

Best For: Sentra

Choose Sentra when stale or contradictory knowledge becomes an operational liability and the people who need that knowledge are split across humans and agents. Operations leads, chiefs of staff, and strategy teams feel this first. A commitment gets made in a Slack thread, a decision reverses in a board deck, and three weeks later an agent restates the deprecated version as current.

Sentra fits when you run cross-functional AI workflows where every agent and every person must read the same facts. What you teach one agent, every agent remembers, because they share one org-wide graph rather than per-entity stores. That shared graph is what lets commitment tracking and contradiction detection work across teams instead of inside a single agent's memory.

The clearest decision signal is reversal risk. If your pricing, scope, or partner commitments change often enough that people quietly act on outdated information, you need a system that knows when a fact stopped being true. Sentra's bi-temporal graph records exactly that, and surfaces what has gone stale before it costs you.

Verdict: Which One to Choose

Zep serves memory to agents through isolated per-entity graphs. Sentra serves one shared graph to your whole company and every agent on it. The right choice follows from which audience you are building for.

Choose Zep when your workload is agent-centric and isolation is a feature, not a limitation. If you run thousands of agents that each need fast, scoped recall, and you want sub-200ms retrieval across millions of separate graphs, Zep's Graphiti substrate is built for exactly that. The per-entity model keeps a user's context out of an agent's context by design, which is what high-throughput agent fleets actually want.

Choose Sentra when stale or contradictory knowledge across people and agents is the real liability. If a commitment made in Slack needs to reach the agent drafting the renewal, or a pricing reversal must invalidate every downstream answer, you need one org-wide graph that humans and agents both read and write. Write-time comprehension against your ontology, commitment tracking, and contradiction detection close gaps that per-entity memory never sees, because no single graph holds both sides of the drift.

The decision rule is short. Building memory for an agent product, pick Zep. Building a brain your ops, strategy, and cross-functional AI workflows all share, pick Sentra. Many enterprises will run both, with Zep behind individual agents and Sentra holding the company-wide layer.

FAQ

Can Zep and Sentra be used together?

Yes. Zep can serve fast per-agent memory while Sentra acts as the org-wide brain that both humans and agents read and write to. Sentra connects through REST or MCP, so it sits as the shared memory layer underneath your existing agents rather than replacing them.

Does Zep support human users, not just agents?

Zep organizes memory into per-entity graphs that include a USER type, but its design centers on serving context to agents. Sentra builds one graph shared by every team member and every agent, so what you teach one agent, every agent and person can recall. The split is audience scope, with Zep optimized for agent workloads and Sentra for shared human-plus-agent context.

How does Sentra's write-time comprehension differ from RAG?

RAG retrieves vectors that are close to a query, which returns what is similar rather than what is correct. Sentra resolves semantics at ingestion against an ontology unique to your organization, so meaning is fixed when a fact enters the graph. That write-time work means agents query a structured graph instead of guessing from a flat embedding store.

Which system handles contradiction detection?

Sentra surfaces contradictions and drift across the whole organization, flagging reversals, stale facts, and unfulfilled commitments. Zep offers Observations that infer patterns within a single graph, such as repeated upgrade behavior, but describes no cross-entity contradiction layer. Sentra owns the cross-entity drift dimension by design.

What compliance certifications does each carry?

Zep holds SOC 2 Type II and offers a HIPAA BAA, with BYOK and BYOC deployment inside your VPC. Sentra holds SOC 2 Type II and ISO 27001, with cloud, isolated VPC, and fully air-gapped on-prem options. Both meet enterprise security bars, and Sentra adds an air-gapped path for the strictest environments.