remem-ai 1.0.0b1

An AI Work Reuse Engine — remember expensive AI work and reuse it intelligently.

Remem

Remember expensive AI work. Reuse it intelligently.

An AI-native work reuse engine for RAG pipelines, AI agents, and LLM applications

Quickstart · API Reference · Architecture · Benchmarks · Roadmap · FAQ

---

Overview

Modern AI applications repeatedly execute expensive operations — vector database searches, LLM calls, tool executions, reranking passes — for requests that are often semantically identical or very similar to ones already served.

Remem sits between your application and its expensive AI operations. It observes what has already been computed, and for each new request it determines the highest level of previous work that can be safely reused — based on semantic similarity, not exact key matching.

User asks: "What is our company's vacation policy?"
User asks: "How many paid leaves do employees get?"   <-- Remem reuses the first answer
User asks: "What are the PTO rules?"                  <-- Remem reuses the first answer

Why Not Just Use Redis?

Redis answers: "Have I seen this exact key before?" Remem answers: "Have I already done similar expensive AI work before?"

	Redis	Remem
Matching strategy	Exact key	Semantic similarity (embeddings)
Scope	Key-value lookup	Full AI pipeline reuse
Deployment	Separate service	Python library — import and go
AI-aware	No	Yes — understands embeddings, models, and prompts

How It Works

Remem makes a three-way decision for every incoming request, based on how similar it is to work that has already been done:

Decision	When it happens	Effect
Full reuse	The request is nearly identical to a past one	The cached response is returned immediately — no vector search, no LLM call
Partial reuse	The request is related, but not identical	Cached retrieval results are reused; only the generation step re-runs
Miss	Nothing sufficiently similar has been seen	The full pipeline runs, and the result is stored for next time

Similarity is only ever compared between compatible executions — Remem is aware of concepts like namespace, knowledge-base version, and model, so updating your data or switching models never results in stale or incorrect reuse.

For the full decision model, thresholds, and configuration options, see the Architecture guide and API Reference.

Features

• Three-level reuse engine — full response reuse, partial retrieval reuse, or a clean miss
• Semantic similarity matching — cosine similarity over embedding vectors, no exact-match required
• Context-aware filtering — namespace, knowledge-base version, prompt version, and model constraints are respected before anything is reused
• Configurable reuse policy — tune thresholds and constraints to match your application
• Durable persistence — a file-backed store that survives process restarts, with atomic writes
• In-memory mode — zero-disk-I/O storage for tests, notebooks, and short-lived jobs
• Pluggable storage — bring your own backend (Redis, Postgres, S3, or anything else)
• Built-in telemetry — hit rate, reuse breakdown, and average similarity, out of the box
• Minimal footprint — a single runtime dependency (numpy)

Installation

Requirements: Python 3.10 or later.

pip install remem-ai

Remem has a single runtime dependency (numpy), which is installed automatically.

Installing from source

git clone https://github.com/harshvardhansingh7/remem.git
cd remem
pip install -e ".[dev]"    # editable install with pytest and ruff

Quickstart

from remem import Client, ExecutionResult

client = Client()   # durable persistence by default

def my_pipeline():
    docs = search_vector_db(query)
    answer = call_llm(query, docs)
    return ExecutionResult(response=answer, references=docs)

outcome = client.get_or_compute(
    query_embedding=embed(query),
    compute_callback=my_pipeline,
)

print(outcome.result)   # cached or freshly computed — Remem decides

New to Remem? Start with the 5-minute Quickstart. For the complete method-by-method reference, see the API Reference.

Documentation

Guide	Description
Quickstart	Get a working integration in under five minutes
API Reference	Every class, method, and configuration option
Architecture	How Remem is designed internally
Benchmarks	What Remem measures and why
Roadmap	Where the project is headed
FAQ	Common questions
Troubleshooting	Fixes for common issues

Contributing

Contributions are welcome. As the project is in beta, architecture discussions and design feedback are especially valuable. Please read CONTRIBUTING.md before opening issues or pull requests.

License

Licensed under the Apache License 2.0.

Author

Harshvardhan Singh — building Remem as an open-source AI infrastructure project to explore distributed systems, storage engines, and high-performance backend engineering.

If this project is useful to you, consider giving it a star and following its progress.

---

⭐ Star the repository if you find Remem useful.