Hebbrix

Hebbrix is a persistent-memory service for AI applications and agents. It stores information across conversations, retrieves relevant memories, and records how facts and relationships change. Its outcome-memory workflow adds explicit feedback about whether an agent's decision helped, rather than relying only on a similar-looking past conversation.

Store and retrieve durable context

The Hebbrix documentation describes memory management, collections, hybrid search and a knowledge graph. Collections organize memories for multi-tenant applications. Retrieval combines semantic, lexical, graph and temporal signals, while versioned memory retains history, provenance and deletion controls.

Applications can call the memory and search APIs directly, or use an OpenAI-compatible chat interface that injects memory into context. A hosted MCP endpoint connects compatible agents, including coding-agent workflows. These are integration choices for a developer to implement; the memory service does not replace the application's own task logic.

Report outcomes and inspect uncertainty

The product page describes an agent reporting delayed success, failure, a business metric or a correction against a decision receipt. Hebbrix updates evidence for the tenant and can use the reported outcomes when ranking memory for subsequent choices. Missing feedback is not treated as a failure.

The API reference distinguishes core resources from advanced or beta functions, including outcome policy evaluation and confidence advice. It explicitly says these features do not authorize actions or establish verified learning superiority. Search scores are ranking signals rather than calibrated correctness probabilities. An integration should therefore inspect no-match and abstention indicators rather than treat the nearest retrieved memory as confirmed evidence.

Smart ingestion is asynchronous by default and returns a job ID; memory becomes searchable when that job completes. That detail matters if your next step immediately searches for newly submitted material. The reference also documents idempotency keys for mutating requests and retry handling for rate-limited responses.

Plans and memory portability

The captured pricing page lists Free with 1,000 monthly credits, Starter at $19 with 25,000, Pro at $99 with 200,000 and Scale at $399 with one million. Pro lists the Outcome Memory API and advanced analytics. Free and Starter pause memory writes at quota exhaustion while reads continue; Pro and Scale can enable overage billing.

Credits measure normalized memory operations, and managed-model calls also report token and estimated-cost usage. Bring-your-own-model-key access is available on supported paid plans. The publisher says memories and the knowledge graph can be exported on any plan, allowing stored data to remain portable when switching the underlying model. Review the relevant feature gates and operation weights before choosing a production plan.