Guide8 min read
Understanding Agent Memory: A Developer's Guide
A deep dive into the four types of memory AI agents need and why each matters for building production-grade systems.
Coming soon
Open Source · Serverless · AWS-native
The memory infrastructure
for AI agents
One API for four memory types. Working, semantic, episodic, and procedural — all serverless, all AWS-native. Give your agents persistent memory in minutes, not months.
DynamoDBpgvectorS3Bedrock Titan
Works withLangGraphLangChainCrewAIAutoGenOpenAI Agents SDK
The problem
AI agents are stateless by default
To give them memory, developers stitch together four databases, four clients, four billing accounts — with no unified query layer.
Four databases to stitch together
Redis for state. Pinecone for vectors. Postgres for structured data. S3 for logs. Four billing accounts, four clients, zero unified memory layer.
No unified query layer
State lives here. Vectors live there. Events are somewhere else. You write glue code instead of writing agents.
Agents forget everything
Without persistence, every invocation starts from scratch. Context, preferences, and history vanish between calls.
The solution
One API. Four memory types.
AWS-native serverless infrastructure. No LLM required for CRUD operations.
<10ms
Working Memory
Key-value state in DynamoDB. Sub-10ms reads with optimistic locking and configurable TTL. Ideal for agent step state and task context.
1024d
Semantic Memory
Natural-language text stored as 1024-dimensional vectors in Aurora pgvector. Auto-embedded via Bedrock Titan. Duplicates are merged, not re-inserted.
Hot + Cold
Episodic Memory
Append-only time-series event log. Hot data in DynamoDB, automatically tiered to S3. Full session replay and time-range queries.
v0.2
Procedural Memory
Tool definitions, prompt templates, schemas, and rules stored in Postgres. Version-controlled and queryable by name. Schema is live; SDK methods ship in v0.2.
Quickstart
Your first agent memory in 15 lines
Install pip install mnemora-sdk and you're ready.
quickstart.py
Python
1from mnemora import MnemoraSync23with MnemoraSync(api_key="mnm_...") as client:4 # Store working-memory state5 client.store_state("agent-1", {"task": "summarize Q4", "step": 1})67 # Semantic memory — auto-embedded server-side8 client.store_memory("agent-1", "User prefers bullet points over prose.")910 # Vector search across all stored memories11 results = client.search_memory("user formatting preferences", agent_id="agent-1")12 for r in results:13 print(r.content, r.similarity_score)1415 # Log an episode to the time-series history16 client.store_episode(agent_id="agent-1", session_id="sess-001",17 type="action", content={"tool": "summarize", "input": "Q4 report"})
Comparison
How Mnemora compares
Concrete data. No hype.
| Feature | Mnemora | Mem0 | Zep | Letta |
|---|---|---|---|---|
| Memory types | 4 (state, semantic, episodic, procedural) | 1 (semantic only) | 2 (semantic + temporal) | 2 (core + archival) |
| Vector search | pgvector 1024d | External DB | Built-in | Built-in |
| LLM required for CRUD | Every op | Every op | ||
| Serverless | ||||
| Self-hostable | Partial | |||
| Multi-tenant | ||||
| LangGraph checkpoints | ||||
| State latency | <10ms | ~500ms | <200ms | ~1s |
Data based on public documentation as of 2025. Subject to change.
Why Mnemora
Designed different by design
No LLM required for CRUD
Mem0 and Letta call an LLM for every memory operation, adding latency and token cost. Mnemora does direct database CRUD. State reads are sub-10ms. No LLM overhead, no rate limits.
Truly serverless
Every component scales to zero when idle. DynamoDB on-demand, Aurora Serverless v2, Lambda, S3. You pay per request. Estimated idle cost: ~$1/month.
Multi-tenant by design
Each API key is scoped to a tenant. Each agent gets an isolated namespace. Data is never mixed at the database layer. Built for SaaS products with multiple end-users.
LangGraph native checkpoints
Drop in MnemoraCheckpointSaver as your LangGraph checkpointer. Each thread_id maps to a Mnemora agent with optimistic locking to prevent concurrent-write data loss.
FAQ
Frequently asked questions
Blog
All posts From the team
Tutorial12 min read
Building Persistent AI Agents with LangGraph and Mnemora
Step-by-step tutorial: how to use MnemoraCheckpointSaver to give your LangGraph agents durable, cross-session memory.
Coming soon
Architecture6 min read
Why Serverless Memory Matters for AI at Scale
The economics of serverless infrastructure for AI workloads — and why stateless-by-default is the wrong architecture choice.
Coming soon
Pricing
Simple, transparent pricing
Start free. Scale as you grow. No surprises.
Free
For exploration and side projects
Free
- 10,000 ops / month
- 1 agent
- 500 MB storage
- Community support
- Working + semantic memory
Starter
For early-stage products
$19/month
- 100,000 ops / month
- 10 agents
- 5 GB storage
- Email support
- All memory types
- API analytics
Most popular
Pro
For production applications
$79/month
- 1M ops / month
- Unlimited agents
- 50 GB storage
- Priority support
- All memory types
- Advanced analytics
- LangGraph checkpoints
Scale
For high-volume teams
$299/month
- 10M ops / month
- Unlimited agents
- 500 GB storage
- Dedicated support + SLA
- All memory types
- Custom integrations
- Invoice billing
All plans include TLS encryption, AWS-native infrastructure, and the full Python SDK. No credit card required for the free tier.
Ready to give your agents memory?
Start in under 5 minutes. No infrastructure to configure. No servers to manage. Just memory that works.