Use Case

AI agents — long-term memory on object storage.

Give your autonomous agents persistent memory, tool retrieval, and per-user context without managing a vector database cluster. Namespaced per user or session, stored in your own bucket, queryable in under 10ms.

Memory patterns

Three types of agent memory ManyVector enables.

Episodic memory

Past conversation recall

Store embeddings of past interactions. Before each turn, retrieve the most relevant memories for the current context. Agents can recall user preferences, prior decisions, and past tasks.

Semantic memory

Knowledge base retrieval

Store facts, documents, and tool descriptions. Agents retrieve relevant knowledge before acting — grounding decisions in verified information rather than hallucinated context.

Tool memory

Dynamic tool selection

Embed tool descriptions and retrieve the most relevant tools for a given task. Scale to hundreds of tools without stuffing all descriptions into the context window.

Namespacing

One namespace per user. Zero cross-contamination.

Create a dedicated namespace per user or per agent session. Memories are isolated — user A cannot retrieve user B's data. Namespaces are lightweight (metadata-only until first write), so creating thousands per day has no overhead.

Namespace branching lets you fork an agent's memory for testing and experimentation — without disrupting the live session.

Namespace per user: complete memory isolation
Namespace per session: ephemeral memory cleared on logout
Shared namespace for global knowledge retrieval
Fork namespaces for eval without disrupting production agents
Storage cost: your bucket, object-storage prices
By the numbers

Long-term memory for autonomous agents.

<10ms

memory recall

memory horizon

Any

agent framework

Multi-agent

shared memory

FAQ

Common questions

How do agents read and write memory with ManyVector?

Use the SDK to upsert memory embeddings on write and query nearest-neighbour memories on recall. Works with LangChain, LlamaIndex, AutoGen, or any agent framework.

How do I prevent memory from growing unboundedly?

Set TTL on memory items, use namespace branching for session-scoped memories, or prune low-relevance memories periodically using a scheduled query.

Can multiple agents share a memory namespace?

Yes. A shared namespace lets agent teams build collective memory. Use metadata filtering to scope queries by agent ID, session, or task.

How does tool retrieval differ from RAG retrieval?

Conceptually identical — embed tool descriptions, query for the most relevant tools given the current task context. ManyVector works the same for both; the difference is what you embed.

Give your agents persistent memory.