MemPalace vs Letta (MemGPT): Verbatim Storage vs Self-Editing Memory in 2026
Two fundamentally different philosophies of AI memory. One stores everything verbatim and lets retrieval do the work. The other lets the agent decide what to remember. Here is how they compare.
MemPalace vs Letta (MemGPT): MemPalace is a free, local-first AI memory system that scores 100% on LongMemEval by storing conversations verbatim with AAAK compression. Letta, evolved from the UC Berkeley MemGPT research project, uses a self-editing tiered memory architecture where the agent manages its own core, archival, and recall memory through tool calls. MemPalace costs $0/year with an MIT license; Letta offers a free tier up to $249/month. Both are open-source, but MemPalace runs fully embedded while Letta requires a server process.
Quick Verdict (TL;DR)
Choose MemPalace if…
- ✓You want verified benchmark accuracy— 100% LongMemEval
- ✓Nothing should be summarized away— verbatim storage
- ✓Zero cost, no server to manage
- ✓You use Claude Code / MCP workflows
Choose Letta if…
- ✓You want agents that manage their own memory
- ✓Tiered memory (core + archival + recall) fits your use case
- ✓You need a full agent framework, not just memory
- ✓Academic roots matter — UC Berkeley research
Feature-by-Feature Comparison
| Feature | MemPalace | Letta (MemGPT) |
|---|---|---|
| LongMemEval Score | 100% (hybrid) / 96.6% (raw) | Not published |
| Pricing | Free (MIT) | $0–249/mo |
| License | MIT | Apache 2.0 |
| Runs Locally | Yes, fully embedded | Self-host (server process) |
| Cloud Option | No (local-first) | Yes (Letta Cloud) |
| Storage Approach | Verbatim + AAAK compression | Agent self-editing (tiered) |
| Memory Architecture | Wings/Rooms/Halls/Closets/Drawers | Core / Archival / Recall |
| Who Decides What to Remember | Retrieval system (store all) | The agent (self-editing) |
| Vector DB | ChromaDB (embedded) | Configurable (Chroma, pgvector) |
| MCP Tools | 19 tools | Agent tool calls |
| Agent Framework | MCP-native tools | Built-in agent runtime |
| Multi-user Support | Per-user palaces | Multi-agent, multi-user |
| Language | Python | Python |
| Origin | Open-source community | UC Berkeley research (MemGPT) |
| Deployment Complexity | pip install + go | Docker / server setup |
Cells highlighted in green indicate the stronger option for that row. Data as of April 2026.
Architecture Comparison
MemPalace — The Memory Palace
MemPalace uses the Memory Palace metaphor: Wings, Rooms, Halls, Closets, and Drawers. Every conversation is stored verbatim, then organized spatially and compressed with AAAK (30x lossless compression).
The system runs a 4-layer retrieval stack(L0–L3): from fast keyword lookup to full semantic reranking. The retrieval system decides what is relevant — the human and AI never have to manually manage what gets stored or forgotten.
Letta — Self-Editing Agent Memory
Letta uses a tiered memory system inspired by operating system memory hierarchies. Core memoryis the agent's active working memory (always in context). Archival memory is a long-term vector store the agent can search. Recall memory stores conversation history for later retrieval.
The key innovation: the agent manages its own memorythrough tool calls. It decides what to write to core memory, what to archive, and what to retrieve — mimicking how a human consciously decides what to remember.
Key philosophical difference: MemPalace stores everything and trusts its retrieval system to surface the right memories at the right time. Letta trusts the agent itself to decide what is worth remembering. MemPalace's approach guarantees nothing is lost; Letta's approach gives the agent more autonomy but risks the agent forgetting something it decided was unimportant. The benchmark results suggest the “store everything” approach yields higher recall accuracy.
Benchmark Deep Dive
LongMemEval (500 questions)
LongMemEval tests how well a memory system can answer questions across long conversation histories. It covers fact recall, temporal reasoning, multi-hop inference, and knowledge updates over time.
MemPalace
100%
Hybrid mode (with Haiku reranking)
96.6%
Raw mode (zero API cost)
Letta (MemGPT)
—
Not published as of April 2026
Why Letta Lacks Published Benchmarks
Letta's architecture is designed around agent autonomy rather than standardized memory recall benchmarks. The agent decides what to store and retrieve, making deterministic benchmarking more complex. The original MemGPT paper focused on demonstrating that agents could manage their own memory effectively, rather than competing on standardized recall metrics.
This does not mean Letta performs poorly — it means we cannot make a direct numerical comparison. MemPalace's published 100% LongMemEval score stands as the only verified number between the two systems.
Methodology note: MemPalace benchmark numbers come from the project's published evaluation suite. Letta has not published comparable LongMemEval results as of April 2026. We will update this page if and when Letta publishes official benchmark scores.
MemPalace is the only system between the two with published LongMemEval scores, achieving 100% in hybrid mode and 96.6% in raw mode. Letta (formerly MemGPT) has not published comparable benchmark results as of April 2026.
Pricing Analysis
MemPalace
$0/year
- ✓MIT license, unlimited use
- ✓All 19 MCP tools included
- ✓Local embedding (no API needed)
- ·Optional: ~$0.001/query for Haiku reranking
Letta (MemGPT)
$0–$249/month
- ·Free: open-source self-host
- ·Developer: free tier (Letta Cloud)
- ·Team: $49/mo per seat
- ·Enterprise: $249/mo + custom pricing
Self-host vs. fully embedded: hidden costs
Both systems can be self-hosted for free, but the operational overhead differs significantly:
- MemPalace:
pip installand go. SQLite + ChromaDB are embedded — no Docker, no server process, no database to manage. - Letta:Self-hosting requires running the Letta server, configuring a database backend, and managing the agent runtime. More powerful, but more operational overhead.
When to Choose MemPalace
- 1
You want verified benchmark accuracy
MemPalace is the only system of the two with published LongMemEval scores — 100% in hybrid mode. If provable recall accuracy matters, the numbers speak for themselves.
- 2
You want true zero-config local memory
pip install, connect to Claude Code, done. No Docker, no server process, no database configuration. Everything is embedded and runs in-process.
- 3
Nothing should be forgotten
MemPalace's verbatim storage means the retrieval system — not the AI — decides what's relevant at query time. No information is ever discarded or summarized away.
- 4
You use Claude Code or MCP workflows
MemPalace was built MCP-native with 19 tools. It integrates directly with Claude Code, Claude Desktop, and any MCP-compatible client out of the box.
- 5
You want zero ongoing costs
MIT license, no subscription tiers, no usage limits. The only optional cost is sub-penny Haiku reranking for maximum accuracy.
When to Choose Letta
- 1
You're building autonomous agents
Letta was designed as an agent framework first, memory system second. If your use case is autonomous agents that need to manage their own context window, Letta's self-editing memory is purpose-built for this.
- 2
You want the agent to control its memory
Letta's core innovation is that the agent decides what to store, update, and retrieve through tool calls. This is a fundamentally different paradigm from MemPalace's store-everything approach — and may be preferable for agents that need editorial control over their own state.
- 3
You need tiered memory hierarchies
Core memory (always in context), archival memory (searchable long-term store), and recall memory (conversation history) map naturally to many agent architectures. If you think in these tiers, Letta will feel intuitive.
- 4
Academic provenance matters
Letta evolved from the MemGPT research paper at UC Berkeley, which introduced the concept of OS-inspired memory management for LLMs. If you value research-backed architecture, Letta has strong academic roots.
- 5
You want a cloud-managed option
Letta Cloud offers a hosted platform so you don't have to run infrastructure. If self-hosting is not an option and you need managed agent memory, Letta has you covered.
Frequently Asked Questions
Is MemPalace really free compared to Letta?+
Yes. MemPalace is MIT-licensed and completely free to use with no restrictions. Letta's open-source version is also free to self-host (Apache 2.0), but Letta Cloud's Team and Enterprise plans cost $49-249/month. MemPalace's only optional cost is approximately $0.001 per query if you enable Haiku reranking for enhanced accuracy.
What is Letta and how does it relate to MemGPT?+
Letta is the commercial evolution of the MemGPT research project from UC Berkeley. The original MemGPT paper introduced the concept of agents that manage their own memory through tool calls, inspired by operating system virtual memory. Letta builds on this with a production-ready platform offering both self-hosted and cloud deployment options.
How does Letta's self-editing memory work?+
Letta gives the AI agent tools to manage three tiers of memory. Core memory is always in the context window — the agent can read and edit it directly. Archival memory is a vector store for long-term facts the agent can search and write to. Recall memory stores conversation history. The agent uses these tools to actively decide what to remember, update, and forget.
Has Letta published LongMemEval benchmark scores?+
As of April 2026, Letta has not published official LongMemEval benchmark scores. MemPalace scores 100% in hybrid mode and 96.6% in raw mode on the same benchmark. Without published numbers from Letta, a direct numerical comparison is not possible.
Can Letta run locally like MemPalace?+
Yes, Letta is open-source and self-hostable. However, self-hosting Letta requires running a server process with its own dependencies and database configuration. MemPalace is designed local-first with embedded SQLite and ChromaDB — no server process needed, everything runs in-process with a simple pip install.
Which system is better for building agents?+
Letta was designed as an agent framework and gives agents direct control over their memory through tool calls. MemPalace was designed as a memory system that integrates with existing agent workflows via MCP. If you're building autonomous agents that need self-editing memory, Letta may be the better fit. If you want to add perfect recall memory to an existing Claude Code or MCP workflow, MemPalace is the natural choice.
Is MemPalace better than Letta for AI memory?+
For verified benchmark accuracy and zero-cost operation, MemPalace has a clear edge — 100% on LongMemEval, fully free, and zero-config local. Letta's advantage is its agent-native architecture where the AI manages its own memory, plus cloud deployment options. Choose MemPalace for maximum recall accuracy and privacy; choose Letta for agent-first architectures where the AI needs autonomous memory control.
Should I switch from Letta/MemGPT to MemPalace?+
Consider switching if you need verified benchmark accuracy (100% on LongMemEval), want zero cloud costs, prefer verbatim storage that preserves full context, or use Claude Code and MCP workflows. Stay with Letta if you're building autonomous agents that need self-editing memory, require the tiered core/archival/recall architecture, or are deeply invested in Letta's agent framework ecosystem.
Ready to try MemPalace?
Get started in under 2 minutes. Install with pip, connect to Claude Code, and give your AI perfect memory — for free.
Last updated: April 8, 2026. Data sourced from official documentation, the MemGPT research paper, and public GitHub repositories.