Letta (formerly MemGPT) pioneered tiered memory for LLM agents. VEKTOR takes a different approach: a persistent local graph that outlives every session. Here's where they diverge.
VEKTOR stores every agent memory as a node in a 4-layer SQLite graph (semantic, causal, temporal, entity). Memories accumulate across all sessions indefinitely — the graph never resets. The AUDN curation layer prevents bloat by evaluating every new input against existing nodes before writing.
The REM cycle runs asynchronously while the agent is idle, compressing 50 raw fragments into 3 distilled insights. Retrieval is always local: 8ms average, no network hop.
Letta's architecture comes from the MemGPT research paper. It divides memory into three tiers: core memory (always in-context), recall memory (searchable conversation history), and archival memory (long-term external storage). An LLM-driven routing layer decides what to retrieve and when.
This tiered model excels for agents that carry out long, complex tasks across many turns — the original MemGPT paper showed a 3.4× improvement on long-horizon benchmarks. The trade-off is latency: every archival retrieval involves an LLM classification step before the actual lookup.
Letta's tiered model is agent-centric — it treats the LLM's context window as the primary interface, routing memory in and out automatically. This works beautifully for deeply stateful agents doing multi-step reasoning.
VEKTOR's model is graph-centric — all memory accumulates in a persistent SQLite graph regardless of what the LLM is doing. Retrieval is explicit: your agent calls memory.recall() and gets back the most relevant context in 8ms. The graph structure (causal, temporal, entity edges) gives you richer retrieval signals than a flat vector search.
Neither is wrong — they solve different problems. Letta shines for autonomous agents with long task horizons. VEKTOR shines for agents that need fast, private, cost-predictable memory in a Node.js MCP-native stack.
VEKTOR is $9/month flat regardless of query volume, memory size, or LLM provider. Letta's cloud tier is usage-based — fine for experimentation, unpredictable at scale. Self-hosting Letta is available but requires managing the infrastructure.
VEKTOR ships a native MCP server — one config line to connect Claude Desktop, Cursor, Windsurf, or VS Code. Letta doesn't currently have a first-party MCP integration. For teams in MCP-native environments, this is the sharpest practical difference between the two.
Letta is the strongest choice if your agent is running complex autonomous workflows in Python and you need the research-proven tiered memory model. VEKTOR is the strongest choice for Node.js MCP-native stacks where speed, privacy, and pricing predictability matter. They rarely compete head-to-head — pick based on your runtime, not the marketing.
Local-first. 8ms recall. MCP-native. $9/month flat.