screenpipe vs Zep (2026)

screenpipe

by Negentropy Labs, Inc.

Zep

by Zep AI

Pros

• ✓Local-first by architecture rather than by policy: capture, storage and search all sit on your own machine, cloud model calls send only the query, and Ollama support means the whole loop can run without a single cloud call.
• ✓A first-party MCP server rather than a claim: npx -y screenpipe-mcp exposes search-content and export-video over a local REST API on port 3030. It installs with one click into Claude Desktop from the app's own settings panel, with documented stdio setups for Claude Code, OpenAI Codex, Cursor and Warp.
• ✓A free plan on one device and two published paid rates: Basic at $21 per month billed annually, and Business at $500 per seat per year, with a seven-day onboarding trial that requires a payment method. Getting running is a desktop install on macOS, Windows or Linux against a documented five-minute quickstart.

Limitations

• –The trust portal lists SOC 2 Type 2, ISO 27001, GDPR and HIPAA as compliant. The vendor's security page describes ISO 27001 as in progress and not yet certified, and GDPR, HIPAA and CCPA as internal programs and assessments rather than third-party certifications. Independent review volume is two Product Hunt reviews and a G2 profile with no reviews yet.
• –Section 4 of the terms has the user consent, by using the service, to the vendor creating Deidentified Data and Analytics Data from usage and using it to improve and train its models. Customer content, output and personal information are excluded without separate written agreement.
• –The public source is under the Screenpipe Commercial License rather than an OSI license: personal, non-commercial and evaluation use can be free, and commercial use requires a paid license. Continuous recording uses roughly 30 GB of local disk per month.

Pros

• ✓Temporal knowledge graph automatically tracks how facts change over time. When information is updated or contradicted, Zep invalidates outdated facts while preserving historical context, making it more accurate than vector-only memory systems for complex multi-session agent workflows.
• ✓Zep publishes retrieval accuracy of 94.7% at 155ms on LoCoMo and 90.2% at 162ms on LongMemEval, its own measurements on two industry benchmarks. Single-shot retrieval avoids the slower agentic tool-call loops used by some memory systems.
• ✓Graphiti MCP server gives Claude Desktop, Cursor, and other MCP clients persistent knowledge graph memory across sessions. Python, TypeScript, and Go SDKs with webhook support and no agent framework lock-in, backed by named enterprise customers including Samsung, Zscaler and HoneyBook.

Limitations

• –Steep pricing jump from free tier (10,000 credits per month for prototyping) to Flex at $125 per month ($104 annual) with no intermediate tier. Early-stage teams that outgrow free credit limits face a significant cost increase with no middle-ground option.
• –Credit-based pricing makes costs difficult to forecast at production scale. Episodes consume 1 credit per 350 bytes, and paid plans auto-purchase more credits whenever the balance falls below 20% - Flex adds 10,000 credits at $25, Flex Plus 40,000 at $75 - so monthly spend is not capped.
• –Self-hosted Graphiti requires Docker, FalkorDB or Neo4j, and LLM API key management. The managed Memory MCP Server removes that setup but is metered in per-account MCP seats that vary by plan, and custom OIDC and UserGroup ABAC require the Enterprise plan.