# Suprflo

> Suprflo is a memory layer for production AI agents. It extracts facts from
> conversations, stores them with bi-temporal structure, and retrieves them via
> hybrid search with cross-encoder reranking. Postgres-native, no vendor lock-in.
> Status: Pre-launch product. Access is currently restricted via waitlist; there 
> is no public pricing or self-serve signup available yet.

Suprflo replaces the practice of stuffing full conversation history into an agent's context
window. It extracts durable facts from each conversation, stores them with both
valid time and transaction time (bi-temporal), and recalls only what a given
query needs. Maintaining a modest 20K-token history costs roughly 85x more per query than retrieval under typical LLM vendor pricing scenarios, and months of continuous usage exceed context limits entirely.

### Primary Use Cases

- **Long-Running AI Agents:** Maintaining evolving user context across infinite interactions.
- **Contextual Customer Support:** Recalling precise historical context from multi-session customer histories.
- **Cross-Session Personalization:** Storing granular preferences without polluting prompt tokens.
- **Historical Auditability:** Applications requiring precise, point-in-time correctness ("what did the agent believe or know on Tuesday?").

### Benchmarks

Measured end-to-end on the open benchmark harness, using identical answerer and judge models, with every failure root-caused:

- **LoCoMo:** 92.1% accuracy on a full 10-conversation set (1,540 questions) on a single held-out pass. Reached 93.0% post-audit verification against source texts.
- **LongMemEval (Provisional):** 93.1% accuracy on a preliminary run of 87 out of 500 questions. Session-level ingestion utilizing top-50 context retrieval.
- **BEAM 100K:** 75.8% pass rate across 20 long conversations (400 rubric-judged questions). Evaluated at an average rubric score of 0.674 with zero ingestion errors.

### Architecture

- **Storage:** Single Postgres 17 instance powered by native `pgvector`.
- **Temporal Model:** True bi-temporal design where facts retain distinct validity windows and are never overwritten.
- **Retrieval Engine:** Hybrid search (Semantic + BM25 + Entity Boosting) with secondary cross-encoder reranking.
- **Memory Subsystems:** Distinct handling for semantic, episodic, and procedural memory types.
- **Decay & Optimization:** Built-in Ebbinghaus forgetting curve decay coupled with an asynchronous reflection engine.
- **Multi-Tenancy:** Strict hierarchical isolation (`Org -> Project -> User -> Agent`) enforced at the SQL `WHERE` clause layer.
- **Data Durability:** Retains raw turn data alongside extracted facts to guarantee zero loss of subtle context.

### Supported Providers

- **Large Language Models (LLMs):** Any OpenAI-compatible endpoint (including GPT-4.1, Claude Sonnet 4, Gemini 2.5 Pro, Groq, Ollama, DeepSeek, and OpenRouter).
- **Embedding Models:** Native support for OpenAI, Cohere, Amazon Bedrock, and Ollama.
- **Vector Infrastructure:** Fully compatible with standard `pgvector` deployments on Postgres 17.
*Note: The integration layer is written as an abstract provider factory, allowing new model backends to be supported in approximately 50 lines of code.*

### FAQ

**What is Suprflo?**
A production memory layer for AI agents. It extracts facts from conversations, deduplicates them, stores them bi-temporally in Postgres, and recalls them via hybrid search.

**How is Suprflo different from mem0?**
While both share the core blueprint of extract-store-retrieve, Suprflo implements a true bi-temporal model, cross-encoder reranking, memory-type differentiation (semantic/episodic/procedural), active Ebbinghaus decay, and an automated reflection engine on vanilla Postgres with standard row-level API exportability.

**Do I have to run my own infrastructure?**
No. Suprflo operates as a fully managed cloud platform. Private enterprise cloud deployments in isolated VPC networks are available upon request.

**How is data isolated in multi-tenant setups?**
Every row maps strictly to `org_id` and `project_id` constraints. All read/write API endpoints filter tenant parameters directly at the database engine level, validated continuously by cross-tenant isolation test suites.

## Essential Resources

- [Homepage](https://suprflo.com/): Product overview, features, and core use-case explanations.
- [Waitlist](https://suprflo.com/waitlist): Early access intake form for the pre-launch developer beta.
- [About](https://suprflo.com/about): Who builds Suprflo - ATIIAD Technologies Pvt. Ltd. (named as Baaz), founded 2018 in Bangalore, and how benchmark numbers on this site are produced.
- [Privacy Policy](https://suprflo.com/privacy): What the site collects, why, and for how long. Analytics via Google Analytics 4 and cookieless PostHog; no ad pixels, no session recording.
- [Terms of Use](https://suprflo.com/terms): Website terms, acceptable use, and how the self-reported benchmark claims should be read.

## Operator Details

- [Baaz](https://baaz.pro): Parent engineering agency and product operator (ATIIAD Technologies Pvt. Ltd.). Contact queries route to mailto:hq@suprflo.com.