Every time you open a new chat with an AI assistant, it starts from nothing. You re-explain your project, your constraints, your vocabulary. The model that helped you debug a gnarly async issue last Tuesday has forgotten the whole thing by Wednesday. If you switch from Claude to ChatGPT mid-project, you start from zero again. This is the problem HeurChain exists to fix.
What HeurChain is
HeurChain is a memory database that sits outside your AI tools and serves all of them. Any agent — ChatGPT, Claude, Cursor, a LangChain pipeline you wrote yourself, an Ollama model running on your laptop — can write memories to your tenant over a plain HTTP call and retrieve them later. One tenant, every model, the same shared history.
The core API is two routes:
POST /store — write a chunk of text into your memory
POST /query — retrieve relevant memories by semantic + keyword similarityThat is it. No SDK required, though we have Python and TypeScript clients if you prefer. If your agent can make an HTTP POST, it works with HeurChain.
What a "tenant" means
Your tenant is a private data silo. It has its own API key, its own isolated vector store, and its own downloadable vault. Nothing written to your tenant is accessible to any other user or any other tenant. HeurChain does not use your stored memory to train models or improve the service. It is a database, not a pipeline into someone else's system.
When you cancel, your data stays exportable for 30 days. After that it is permanently deleted. That policy is in the Terms of Service, not just the marketing.
Why we built it
We build AI-powered tools ourselves. Over the past year we accumulated five or six AI tools we use daily — ChatGPT for drafting, Claude for analysis, Cursor for code, a custom LangChain agent for research tasks, Hermes (our in-house model) for memory consolidation. Each one had its own notion of "context," and none of them could see what the others had learned.
We tried the obvious patches: custom system prompts, Claude Projects, ChatGPT memory, copy-pasting context between sessions. All of them degraded over time or broke when we switched tools. We wanted something durable, queryable, and portable across every model we touch.
We also wanted something we actually owned. ChatGPT's memory lives inside OpenAI's system. Anthropic's conversation history lives inside Anthropic's system. Neither is queryable, neither is exportable in a useful format, and neither is readable from the other. HeurChain's memory is a database in the literal sense — a store you can query, back up, migrate, or delete on your schedule.
The retrieval approach
HeurChain uses a hybrid retrieval strategy: dense vector search (bge-base-en-v1.5 embeddings) combined with BM25 keyword search, fused via Reciprocal Rank Fusion. This matters in practice because neither pure vector search nor pure keyword search is good enough on its own.
Vector search finds conceptually similar content even when the exact words don't match. Keyword search handles proper nouns, identifiers, and domain-specific terms that vector similarity tends to blur. RRF combines both lists into a single ranked result without needing to tune a weighting coefficient. We wrote a separate post on why this matters for agent memory retrieval specifically.
Where it runs
HeurChain hosted runs on Hetzner infrastructure in Falkenstein, Germany, with Cloudflare in front. Data is stored in the EU. The full stack — broker, Qdrant for vector storage, Redis for working memory, the embedding service — is MIT-licensed open source. You can run it yourself on a single VPS in about ten minutes if you prefer.
What we are shipping now
The hosted service is live at $5/month for a solo tenant and $49/month for a team. Both plans include unlimited agents. You do not pay per AI tool you connect. The whole point is that every tool you use writes into one shared store — charging per agent would defeat the purpose.
If you want to see the code before committing, the GitHub repo is public. If you want to try the hosted service, go to heurchain.com/pricing. We are a small team and we read every support email.