MemGPT (now Letta) pioneered early research into OS-like memory management for LLMs (working context vs archival memory). However, MemGPT is built as a Python application wrapper rather than a dedicated database engine.
1. Application Framework vs. Native Database Engine
MemGPT manages memory by injecting explicit prompt instructions into the LLM system prompt (*"System: You have working memory slot A..."*) and storing records in standard PostgreSQL or SQLite tables.
This creates two major bottlenecks:
- Enormous Token Overhead & Latency: Injecting memory slots into system prompts consumes thousands of context tokens per turn, driving latency above 350ms per interaction.
- Prompt Injection Vulnerability & Drift: LLMs occasionally ignore or overwrite system prompt memory rules under complex multi-turn chats.
Axonio takes the opposite approach: memory is managed by a compiled Rust/C++ database engine that sits outside the LLM prompt loop, executing vector search and graph traversal in sub-millisecond local speed.
2. Architectural Differences: Latency & Memory Footprint
| Dimension | MemGPT / Letta | Axonio Database Engine |
|---|---|---|
| Language & Runtime | Python Framework | Native C++20 / Rust Engine |
| P99 Query Overhead | 350ms - 1,200ms (LLM prompt loop) | 0.38ms (Native buffer search) |
| Memory State Isolation | Prompt text slots | Git-style state branching |
| Contradiction Resolver | Manual overwrite | Background Reflection Engine |