Heartwood vs Letta — provable memory vs the stateful-agent framework
Naming: “Letta” is the current framework name; “MemGPT” now refers to the original research/design pattern. Use “Letta”; “formerly MemGPT” appears once for search.
TL;DR
Letta is a platform for building and operating stateful agents with self-editing, persistent memory blocks (the framework formerly known as MemGPT). Heartwood is a governed memory store: signed records, policy-before-ranking, and a re-executable key-destruction proof. They solve adjacent problems — Letta runs the agent; Heartwood governs the memory.
At-a-glance table
| Axis | Heartwood Memory | Letta |
|---|---|---|
| Governance granularity | Per individual memory record | Not evaluated in public docs (pricing lists RBAC; detailed RBAC page currently 404s) |
| Provenance signing | Per-record Ed25519, fail-closed read | Not evaluated in public docs |
| Tamper-evident audit | Hash-chained audit log | Not evaluated in public docs |
| Policy-before-ranking | Yes | Not evaluated in public docs (core blocks are always-visible) |
| Erasure / RTBF | Crypto-shred + key-destruction proof | Block/archival-passage/agent deletion APIs (not cryptographic or verified RTBF) |
| Interface | Python library + governed MCP server | TypeScript/Python SDKs and REST API; MCP/external DBs as tools |
| Deployment | Self-hosted, embedded | Open-source self-hosted Docker/server, or managed Letta cloud |
| License | Source-available (BSL 1.1); 0.1.x MIT | Apache-2.0 (Letta repo); managed service terms separate |
| Pricing | Free / Team $349·mo / Pro $6,000·yr (early access) | Free $0 / Pro $20·mo / API Plan $20·mo + usage / Enterprise custom |
| Best-for | Record-level, re-executable proof | Building/operating stateful agents with self-editing memory |
“Not evaluated in public docs” = not found in Letta's current primary documentation as of 2026-07-15; it is not a claim the feature is absent. Vendor pricing reverified immediately before publish.
Comparison by dimension
Different layer
Letta is an agent framework/runtime; Heartwood is a governed memory store. The honest comparison is “run the agent” vs “govern its memory” — they can compose.
Deletion & governance
Letta documents block, archival-passage, and agent deletion APIs — but not cryptographic or verified RTBF, and its detailed RBAC page currently 404s, so governance granularity couldn't be verified. Heartwood adds a per-subject key-destruction proof and per-record policy.
Who Letta is best for
Teams building and operating stateful agents with self-editing, persistent memory.
Who Heartwood is best for
Teams that need a governed, provable memory layer under whatever agent framework they run.
Can Heartwood run underneath Letta?
As the governed memory store behind the agent's memory calls (shape-compatible example contract, not an official integration).
See the governed-memory modelFAQ
Is Letta the same as MemGPT?
Letta is the current framework name; MemGPT names the original research pattern.
Is Letta open source?
The Letta repository is Apache-2.0; managed service terms are separate.