thinkingmemory / cloud
the memory database for AI agents

Managed recall for your agents.

Point your agent at one endpoint: it stores experience and gets back the right context, hybrid-retrieved, reranked, and packed to a token budget. Lifecycle, bitemporal audit, and per-tenant isolation, fully managed.

Why a memory database, not a vector store

Recall, not search

Intent in, the right context out: cited and packed to your token budget. No vector wrangling, no re-ranking glue, no prompt stuffing.

Self-curating memory

Decay, consolidation, supersession, and contradiction resolution run continuously, so useful memories rise and stale ones fade.

Hybrid retrieval

Vector similarity + full-text keyword + recency, fused with Reciprocal Rank Fusion and reranked with a cross-encoder.

Bitemporal + audit

Query what an agent believed at any point in time, trace provenance, and review a full audit log.

Multi-tenant by design

Per-tenant isolation enforced at the database with row-level security and hash partitioning.

Open core + MCP

Apache-2.0 engine you can self-host, server-side embeddings (send text, not vectors), and an MCP-native interface.

How it works

1

Remember

POST experience to one endpoint: episodic, semantic, procedural, or working.

2

Recall

Send an intent and a token budget; get back deduped, cited context, ready to drop into the prompt.

3

It self-curates

The lifecycle engine decays, consolidates, and forgets in the background, and recall sharpens as your agent runs.

Frequently asked questions

Is ThinkingMemory a vector database?

No. A vector store returns nearest neighbours; ThinkingMemory returns the useful context for what your agent is doing right now, ranked and packed to a token budget with citations. Vectors are one signal among several.

Do I have to send embeddings?

No. You send text. Embeddings are generated server-side with a local model, so there are no per-token embedding fees and nothing extra to run.

How is recall different from RAG?

RAG usually means 'retrieve similar chunks and stuff the prompt'. Recall fuses vector, keyword, recency, and graph signals, reranks them, then packs a budgeted, cited context, and it learns from what it surfaces.

Can I use it with any agent framework or LLM?

Yes. ThinkingMemory is agent-agnostic and works with any framework or model. It's also MCP-native, so it drops into MCP-compatible clients directly.

Is it open source?

The core engine is open source under Apache-2.0. ThinkingMemory Cloud is the managed, multi-tenant service built on top of it.

How is my data isolated?

Every memory is scoped by tenant and enforced at the database with PostgreSQL row-level security plus hash partitioning, in addition to application-level scoping.

Give your agents a memory.

Start free