Skip to content
Comparison

Korva vs MemGPT / Letta

MemGPT and Letta are general-purpose memory frameworks for LLM agents. Korva is purpose-built for engineering teams: persistent code-level memory, commit-time guardrails, MCP-native, and local-first. Here is when each fits.

Different problems, different shapes

MemGPT (now Letta) is a research-grade framework for giving LLM agents persistent context across conversations. It is genuinely good at what it does — and it is not aimed at engineering teams. There is no IDE integration, no commit-time enforcement, no curated stack knowledge.

Korva starts from a narrower problem: "my engineering team uses Claude Code / Cursor / Copilot every day, and I want them to actually remember our architecture." That narrowness pays off — every component (Vault, Sentinel, Lore, Forge, Beacon, Hive) is opinionated for code, commits, IDEs, and teams.

Capability Korva MemGPT / Letta
Primary purpose Persistent memory + commit-time guardrails for engineering teams using AI assistants General-purpose long-term memory framework for any LLM agent
Storage Local SQLite per machine + FTS5 full-text search; optional Hive sync Postgres + pgvector (Letta) / SQLite (MemGPT) — typically remote-hosted
IDE / editor integration 8 IDEs out of the box — Claude Code, Cursor, VS Code, Copilot, Windsurf, Codex, OpenCode, Gemini None native — must wire up your own client
Architecture enforcement Sentinel: 21 commit-time rules (HEX/SEC/NAM/ARC/QC/DEPS), extensible via YAML Not the goal — agents may store rules but nothing enforces them
Curated stack knowledge 25 Lore scrolls covering NestJS, React/Next, Stripe, Docker/K8s, MCP, security, etc. No curated knowledge layer — bring your own
Workflow structure 5-phase Spec-Driven Development with human approval gates, persisted in vault Free-form agent loop
Privacy default Local-first by default. Cloud sync is opt-in and content-addressed. Cloud-leaning; self-host requires infra setup
License model MIT core; single commercial Teams tier ($9/u/mo launch, $19 regular) Apache 2.0 (Letta) — fully open source
When to choose this Engineering team using Claude Code/Cursor/Copilot every day Building a generic memory-augmented chatbot or research agent

When to choose Letta instead

  • You're building an agent that lives outside an IDE — a customer-support bot, a research assistant, a long-running RPA task.
  • You need a generic memory abstraction with rich tooling for hierarchical context, sleep-time compaction, and message-bus integrations.
  • You want a fully-Apache-licensed, hosted memory service with a UI for non-engineers.

For those use cases Korva is the wrong shape — go use Letta. For an engineering team, Korva is purpose-built and will be in production faster.

Built for code, commits, and teams

Install in 30 seconds. Works with the editor your team already uses.

Install Read the docs