thinqOS, Zep and Graphiti, compared.

The short answer

As of September 22, 2026. Products change; we re-check this page regularly.

Graphiti is ahead of thinqOS on questions about time, such as what was true in March and when that changed, and on searching several steps across a graph. Zep Cloud adds a governance and compliance layer that thinqOS does not match. thinqOS is ahead on the state of each belief: how confident it is, how much it still matters, who may see it, and keeping disagreement instead of overwriting it.

Three things share the Zep name

Graphiti is an open-source Python library under the Apache 2.0 license. You bring a graph database (Neo4j, FalkorDB, Kuzu or Amazon Neptune) and a language model. We read its code at the commit dated September 21, 2026.

Zep Cloud is Zep's hosted, closed-source service, which Zep describes as built on Graphiti. Its server code is not public, so what we say about Zep Cloud comes from Zep's documentation, and the table marks it that way.

Zep Community Edition, the older self-hosted server, is deprecated and unsupported, and its code now sits in a legacy folder. This page does not compare against it.

In Graphiti, each fact is an edge between two entities with four timestamps: when it became true, when it stopped being true, when the system learned it, and when the system learned it had stopped. When a new fact contradicts an older one, a language model call marks the older one invalid, and nothing is deleted. thinqOS records when a belief was created, reinforced or challenged, but not when it was true in the world.

Who should choose Zep or Graphiti instead

We would point you to Graphiti or Zep in these cases.

You are building your own agent and want a library you control: your database, your models, prompts you can read. That is Graphiti. thinqOS is not open source. It can be self-hosted from a signed release, with an install token we issue for each installation.

Your questions are about facts over time, such as which plan a customer was on last quarter. Graphiti and Zep Cloud model that directly, and thinqOS has no model of when a belief was true.

You serve many end users and face a security review. That is Zep Cloud. Its documentation lists SOC 2 Type II, a HIPAA business associate agreement on Enterprise, role-based and attribute-based access control, single sign-on, audit and API logs, and bring-your-own-key and bring-your-own-cloud deployments. thinqOS has no compliance attestation.

You need multi-hop graph search, or your own entity and relationship types.

You want a free start. Zep's pricing page, read September 22, 2026, lists a free plan with 10,000 credits a month, Flex at $125 a month with 50,000 credits, Flex Plus at $375 a month with 200,000 credits, and a custom Enterprise plan. One credit covers an episode of up to 350 bytes. Graphiti itself costs only your own infrastructure and model calls.

Where Zep and Graphiti are ahead

Time. Graphiti keeps both valid time and system time on every fact, and handles facts that arrive out of order. In our test without model calls, an older fact that arrived late did not overwrite a newer one.

Graph search. Graphiti combines keyword search, vector search and graph traversal up to three steps, with five rerankers. thinqOS belief retrieval uses vector search with its own ranking, weighted by relevance, confidence, salience and recency, and does not follow relations at recall time at all. Relations are recorded and used when beliefs are consolidated or found to conflict. Keyword search in thinqOS covers conversations and documents, not beliefs.

Matching names. Graphiti matches entity names exactly, then by fuzzy similarity, then with a model call. thinqOS matches entities by exact name and known aliases, with no fuzzy step.

Your own schema. Graphiti and Zep let a developer define entity and relationship types. thinqOS uses a fixed set of entity types. The predicate that links two entities is free text the model generates, so it is neither a fixed taxonomy nor one a developer defines.

Learning from verified outcomes. Zep's documentation describes Agent Skills compiled from agent runs that recorded an outcome. thinqOS stores procedures and lessons, but its learning from outcomes is partial.

Two caveats from Graphiti's code. Its default search returns facts marked invalid alongside current ones unless the caller filters them out. And whether a new fact correctly invalidates an old one rests on one small-model judgment. Graphiti also sends anonymous telemetry by default, and we found no authentication in the code of its self-run MCP server.

Where thinqOS is ahead

Each point below is in thinqOS's code and running today.

Confidence on each belief. Repeated evidence is blended into one confidence value, and a stated belief outranks an inferred one whichever arrived first. Graphiti has no confidence field.

Salience and forgetting. Salience is tracked apart from confidence, rises when a belief is used, and decays over time. Protected beliefs have a floor, and unprotected ones that fall low enough are archived, which removes them from recall without deleting them. Graphiti has no salience or decay, and Zep's documentation describes none.

Consolidation with a cap. Recurring related beliefs can be lifted into a summary belief whose confidence cannot exceed 0.85 or the average of its sources. It stays unconfirmed until confirmed and keeps a link to each source. Zep Cloud's Observations also derive patterns from recurring facts, per its documentation; the cap and the source links are what differ.

Disagreement kept. Contradicting beliefs stay linked, and a belief can be retracted with a reason without being deleted. Graphiti resolves a contradiction by marking the older fact invalid. For addresses and account state that is usually right. For judgments and preferences, keeping both avoids silently picking a winner.

Audience per belief. Each belief can carry its own audience, which is finer than access control over a whole graph. It is not a substitute for Zep's compliance layer.

Budgeted recall. Recall has token and time budgets for each tier, and ranks by relevance, confidence, salience, recency and goals.

Capture where people already work. Hooks for Claude Code, Codex, Grok Build and VS Code send work into the Mind; with Zep and Graphiti, the developer's application has to send the data.

Open export. A Mind exports as JSON-LD, YAML-LD or a Markdown vault, as well as a signed archive. We found no bulk export in Zep's v3 documentation index, though one may exist.

What is partial in thinqOS

thinqOS checks for contradictions when a belief is confirmed, stated directly or resolved, not on every automatic extraction. Its event log does not yet record every kind of change. Choosing what reaches the context window is still being built, goal pursuit is largely unbuilt, and learning from outcomes is partial. We found no dedicated audit log of access.

Choose thinqOS when

Choose thinqOS when you want a persistent Mind for a person or an agent that holds beliefs with confidence, lets what stops mattering fade, keeps disagreement visible, and follows that person across Claude Code, Codex and the other tools it connects to, rather than a memory backend wired into one application.