Making machines that learn. Create stateful agents that remember everything, learn continuously, and improve themselves over time.
Based on the limited social mentions available, users appear to have concerns about **Letta's performance at scale**, particularly citing **200-500ms latency per operation** due to routing all memory operations through an LLM. This latency issue is seen as a significant enough problem that developers are building alternative solutions specifically to avoid it. The multiple YouTube mentions suggest there's active interest and discussion around Letta in the AI agent community. However, without detailed reviews or pricing information, it's difficult to assess overall user satisfaction or cost concerns beyond the core performance complaint.
Mentions (30d)
0
Reviews
0
Platforms
2
GitHub Stars
21,824
2,303 forks
Based on the limited social mentions available, users appear to have concerns about **Letta's performance at scale**, particularly citing **200-500ms latency per operation** due to routing all memory operations through an LLM. This latency issue is seen as a significant enough problem that developers are building alternative solutions specifically to avoid it. The multiple YouTube mentions suggest there's active interest and discussion around Letta in the AI agent community. However, without detailed reviews or pricing information, it's difficult to assess overall user satisfaction or cost concerns beyond the core performance complaint.
Features
Use Cases
Industry
research
Employees
18
Funding Stage
Seed
Total Funding
$10.0M
754
GitHub followers
47
GitHub repos
21,824
GitHub stars
20
npm packages
Show HN: Mnemora – Serverless memory DB for AI agents (no LLM in your CRUD path)
Hi HN,<p>I built Mnemora because every AI agent memory solution I evaluated (Mem0, Zep, Letta) routes data through an LLM on every read and write. At scale, that means 200-500ms latency per operation, token costs on your memory layer, and a runtime dependency you don't control.<p>Mnemora takes the opposite approach: direct database CRUD. State reads hit DynamoDB at sub-10ms. Semantic search uses pgvector with Bedrock Titan embeddings — the LLM only runs at write time to generate the embedding vector. All reads are pure database queries.<p>Four memory types, one API: 1. Working memory: key-value state in DynamoDB (sub-10ms reads) 2. Semantic memory: vector-searchable facts in Aurora pgvector 3. Episodic memory: time-stamped event logs in S3 + DynamoDB 4. Procedural memory: rules and tool definitions (coming v0.2)<p>Architecture: fully serverless on AWS — Aurora Serverless v2, DynamoDB on-demand, Lambda, S3. Idles at ~$1/month, scales per-request. Multi-tenant by default: each API key maps to an isolated namespace at the database layer.<p>What I'd love feedback on: 1. Is the "no LLM in CRUD path" differentiator clear and compelling? 2. Would you use this over Mem0/Zep for production agents? What's missing? 3. What memory patterns are you solving that don't fit these 4 types?<p>Happy to answer architecture questions.<p>SDK: pythonpip install mnemora<p>from mnemora import MnemoraSync<p>client = MnemoraSync(api_key="mnm_...") client.store_memory("my-agent", "User prefers bullet points over prose") results = client.search_memory("output format preferences", agent_id="my-agent") # [0.54] User prefers bullet points over prose Drop-in LangGraph CheckpointSaver, plus LangChain and CrewAI integrations.<p>Links: 5-min quickstart: <a href="https://mnemora.dev/docs/quickstart" rel="nofollow">https://mnemora.dev/docs/quickstart</a> GitHub: <a href="https://github.com/mnemora-db/mnemora" rel="nofollow">https://github.com/mnemora-db/mnemora</a> PyPI: <a href="https://pypi.org/project/mnemora/" rel="nofollow">https://pypi.org/project/mnemora/</a> Architecture deep-dive: <a href="https://mnemora.dev/blog/serverless-memory-architecture-for-ai-agents" rel="nofollow">https://mnemora.dev/blog/serverless-memory-architecture-for-...</a>
View originalPricing found: $20 / month, $100 / month, $200 / month, $20 / month, $0.10 / active
Show HN: Mnemora – Serverless memory DB for AI agents (no LLM in your CRUD path)
Hi HN,<p>I built Mnemora because every AI agent memory solution I evaluated (Mem0, Zep, Letta) routes data through an LLM on every read and write. At scale, that means 200-500ms latency per operation, token costs on your memory layer, and a runtime dependency you don't control.<p>Mnemora takes the opposite approach: direct database CRUD. State reads hit DynamoDB at sub-10ms. Semantic search uses pgvector with Bedrock Titan embeddings — the LLM only runs at write time to generate the embedding vector. All reads are pure database queries.<p>Four memory types, one API: 1. Working memory: key-value state in DynamoDB (sub-10ms reads) 2. Semantic memory: vector-searchable facts in Aurora pgvector 3. Episodic memory: time-stamped event logs in S3 + DynamoDB 4. Procedural memory: rules and tool definitions (coming v0.2)<p>Architecture: fully serverless on AWS — Aurora Serverless v2, DynamoDB on-demand, Lambda, S3. Idles at ~$1/month, scales per-request. Multi-tenant by default: each API key maps to an isolated namespace at the database layer.<p>What I'd love feedback on: 1. Is the "no LLM in CRUD path" differentiator clear and compelling? 2. Would you use this over Mem0/Zep for production agents? What's missing? 3. What memory patterns are you solving that don't fit these 4 types?<p>Happy to answer architecture questions.<p>SDK: pythonpip install mnemora<p>from mnemora import MnemoraSync<p>client = MnemoraSync(api_key="mnm_...") client.store_memory("my-agent", "User prefers bullet points over prose") results = client.search_memory("output format preferences", agent_id="my-agent") # [0.54] User prefers bullet points over prose Drop-in LangGraph CheckpointSaver, plus LangChain and CrewAI integrations.<p>Links: 5-min quickstart: <a href="https://mnemora.dev/docs/quickstart" rel="nofollow">https://mnemora.dev/docs/quickstart</a> GitHub: <a href="https://github.com/mnemora-db/mnemora" rel="nofollow">https://github.com/mnemora-db/mnemora</a> PyPI: <a href="https://pypi.org/project/mnemora/" rel="nofollow">https://pypi.org/project/mnemora/</a> Architecture deep-dive: <a href="https://mnemora.dev/blog/serverless-memory-architecture-for-ai-agents" rel="nofollow">https://mnemora.dev/blog/serverless-memory-architecture-for-...</a>
View originalRepository Audit Available
Deep analysis of letta-ai/letta — architecture, costs, security, dependencies & more
Yes, Letta offers a free tier. Pricing found: $20 / month, $100 / month, $200 / month, $20 / month, $0.10 / active
Key features include: Persistent agents instead of stateless sessions, Always improving and learning, Own your memory and port it across models, Chat from any device, run on any environment, Compatible with your API keys and coding plans, Install the desktop app, Use in the terminal, Build with the SDK.
Letta is commonly used for: Persistent agents instead of stateless sessions.
Letta has a public GitHub repository with 21,824 stars.
Based on user reviews and social mentions, the most common pain points are: token cost.
