The Explosive Growth of Open-Source Memory Layers (And Why It Matters)

The growth numbers (primary sources)

Mem0

• Launched January 2024.
• 41,000+ GitHub stars (as of October 2025 announcement).
• 13M+ Python package downloads.
• 35M API calls in Q1 2025, 186M in Q3 2025. ~30% month-over-month growth.
• 80,000+ signed-up developers.
• $24M total raised ($3.9M seed + $20M Series A). Led by Basis Set Ventures; YC, Peak XV, GitHub Fund, Kindred Ventures.
• Exclusive memory provider for AWS Agent SDK.
• (TechCrunch)

Zep / Graphiti

• Graphiti engine: 24K+ GitHub stars.
• Zep platform rebranded to "context engineering platform" in 2026.
• Open-source Apache 2.0 for Graphiti; Zep platform is cloud-only.
• Benchmark lead on LongMemEval (63.8% GPT-4o) vs Mem0's 49.0%.

Letta (formerly MemGPT)

• $10M seed, Felicis Ventures (September 2024).
• Apache 2.0 license.
• OS-style tiered memory architecture.

Supermemory

• Founded by a 19-year-old.
• Backed by Google executives.
• MCP-first, optimized for Claude Code and coding agents.

Why the growth is happening

Three structural forces drive the open-source memory-layer growth:

1. 01 The market is converging on memory as a distinct category. Taranjeet Singh, Mem0 CEO: "Every agentic application needs memory, just as every application needs a database." This is not a product claim - it’s a category claim, and the category is accepting it.
2. 02 Stateless LLM to stateful application shift. Users now expect agents to remember. ChatGPT shipped user-facing memory in 2024. The competitive floor moved.
3. 03 Open-source is the trust play. After the MemPalace / GBrain / Delve benchmark-fraud incidents, open-source is the only credible claim. Closed benchmarks from a closed vendor lost credibility.

What’s driving the specific architectural shifts

Every mature open-source memory layer in 2026 has added or is adding:

• Graph extensions on top of vector storage. Mem0 added Mem0g (Neo4j optional). Zep was graph-first from launch. Cognee is graph-first. Even LangChain’s LangMem is adding graph primitives.
• Temporal validity - tracking when facts become invalid, not just when they were recorded.
• Cross-session persistence - facts that survive beyond a single conversation.
• Conflict resolution - updating instead of just appending when new data contradicts old.
• MCP support - making the memory layer callable from Claude Desktop, Claude Code, Cursor, Windsurf.

The architectural floor is rising across the whole market, not just at the leaders.

Where the consolidation is going

Three predictions for 2026-2027:

1. 01 Mem0 becomes the default volume player. The AWS deal is decisive. Once the cloud incumbent picks a memory layer, the developer default moves with it.
2. 02 Zep / Graphiti wins temporal reasoning and enterprise. The benchmark lead on LongMemEval + the context-engineering platform rebrand position it for agents where validity windows matter - CRM, compliance, account management.
3. 03 A new category emerges above storage. Reactive memory (store + retrieve) is commoditizing. The next category is proactive memory - systems that continuously reason over stored context and push recommendations to agents. This is where Genios sits. The category is not yet named. It will be, within 18 months.

What this means for your stack in 2026

If you are picking a memory layer now:

• For personalization + broad compatibility: Mem0.
• For temporal reasoning + enterprise: Zep / Graphiti.
• For long-running agents that need OS-level memory: Letta.
• For coding agents on MCP: Supermemory.
• For multi-agent fleets that need proactive context instead of reactive retrieval: Genios.

If you are picking a memory layer for 2027: consolidation will have reduced the field. The survivors will be the ones that ship enterprise features (SOC 2, HIPAA BAA, on-prem) and prove reproducible benchmarks.