remem-ai 0.6.0

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

Remem

Remember expensive AI work. Reuse it intelligently.

An AI Work Reuse Engine for Retrieval-Augmented Generation (RAG), AI Agents, and LLM Applications.

---

Installation

pip install remem

Remem has a single runtime dependency (numpy) and works on Python 3.10+.

Quickstart

Integrating Remem takes a few lines. Wrap any expensive AI step in get_or_compute: pass the query embedding and a callback that does the real work. Remem decides whether to reuse a previous execution or run the callback, then remembers the result for next time.

from remem import Client, ExecutionResult

client = Client()  # sensible defaults; durable JSON persistence out of the box

def expensive_pipeline():
    docs = search_vector_db(query)        # your retrieval
    answer = call_llm(query, docs)        # your LLM call
    return ExecutionResult(response=answer, references=docs)

outcome = client.get_or_compute(
    query_embedding=embed(query),         # your embedding model
    compute_callback=expensive_pipeline,
)

print(outcome.decision)   # RESPONSE_REUSED | RETRIEVAL_USED | MISS
print(outcome.result)     # the answer (reused or freshly computed)

Need ephemeral storage (tests, notebooks)? Swap the backend without touching anything else:

from remem import Client, InMemoryStorage

client = Client(storage_backend=InMemoryStorage())  # nothing written to disk

Gradually adopt the advanced features when you need them — execution contexts, custom policies, and metadata matching:

from remem import Client, ExecutionContext, ReusePolicy

client = Client(
    policy=ReusePolicy(
        retrieval_threshold=0.80,   # reuse retrieved docs above this similarity
        response_threshold=0.95,    # reuse the whole LLM response above this
        require_same_model=True,    # never reuse across different models
    )
)

context = ExecutionContext(namespace="hr-bot", kb_version="2024.1", model="gpt-4o")
outcome = client.get_or_compute(embed(query), expensive_pipeline, context=context)

A complete, runnable RAG example lives in examples/rag_reuse.py.

---

Why Remem?

Modern AI applications repeatedly execute expensive operations for requests that are often semantically similar.

A typical AI workflow may involve:

• Generating embeddings
• Searching vector databases
• Retrieving knowledge chunks
• Reranking retrieved documents
• Constructing prompts
• Calling Large Language Models (LLMs)
• Executing tools or SQL queries
• Running multi-step agent workflows

Although user queries may be phrased differently, they frequently require nearly identical computation.

For example:

"What is our company's vacation policy?"

"What are the PTO rules?"

"How many paid leaves do employees receive?"

Most AI systems execute the entire pipeline independently for each request, even when much of the previous work could be safely reused.

This leads to:

• Higher inference costs
• Increased latency
• Unnecessary vector searches
• Repeated reranking
• Duplicate LLM calls
• Repeated tool execution
• Wasted infrastructure resources

Remem exists to eliminate redundant AI work.

Instead of treating every request as completely new, Remem remembers previously executed work and determines whether any part of it can be safely reused.

---

Vision

Remem is an AI Work Reuse Engine.

Rather than replacing your existing infrastructure (Redis, Postgres, Pinecone, Qdrant, Weaviate, Milvus, pgvector, or custom retrieval systems), Remem integrates alongside your application and continuously observes expensive AI operations.

When a new request arrives, Remem analyzes previously completed work and determines the highest level of computation that can be safely reused.

Depending on the request, this may include:

• Previously generated LLM responses
• Retrieved knowledge chunks
• Reranked search results
• Tool outputs
• SQL query results
• Agent execution artifacts
• Other reusable intermediate computations

If no reusable work exists, the application executes normally, and Remem learns from the new execution for future requests.

The long-term vision is to become a lightweight infrastructure component that AI engineers can integrate into existing RAG systems, AI agents, and LLM applications with minimal changes.

---

The Problem

A typical Retrieval-Augmented Generation (RAG) pipeline looks like this:

                 User Query
                      │
                      ▼
            Generate Embedding
                      │
                      ▼
          Search Vector Database
                      │
                      ▼
            Retrieve Knowledge
                      │
                      ▼
                Rerank Results
                      │
                      ▼
            Construct Prompt
                      │
                      ▼
                 Call the LLM
                      │
                      ▼
                  Final Answer

Even when multiple users ask nearly identical questions, this entire pipeline is often executed repeatedly.

As applications scale, this repeated computation becomes one of the largest contributors to latency and infrastructure cost.

---

How Remem Works

Instead of assuming every request requires a full execution, Remem attempts to reuse previous work whenever it is safe to do so.

                 Incoming Request
                        │
                        ▼
              Generate Embedding
                        │
                        ▼
               Semantic Similarity
                        │
                        ▼
            AI Work Reuse Decision
                        │
     ┌──────────────────┼──────────────────┐
     │                  │                  │
     ▼                  ▼                  ▼
Reuse Response   Reuse Retrieval     Execute Pipeline
     │                  │                  │
     │                  ▼                  ▼
     │           Skip Retrieval      Store New Execution
     │             & Reranking             │
     └──────────────────┴──────────────────┘
                        │
                        ▼
                 Return Result

Rather than acting as a traditional cache, Remem behaves like a decision engine that selects the highest level of reusable computation for each request.

---

What Can Remem Reuse?

Depending on the request and available metadata, Remem may reuse:

• Entire LLM responses
• Retrieved knowledge chunks
• Reranked search results
• Prompt construction artifacts
• Tool execution results
• SQL query results
• Agent execution artifacts
• Future AI workflow outputs

Not every request can safely reuse every artifact.

For example:

• If the knowledge base has changed, Remem may skip response reuse but still reuse retrieval results.
• If the retrieval results have changed, Remem executes a fresh retrieval.
• If no previous work is reusable, the application executes normally.

Rather than guaranteeing that every request avoids an LLM call, Remem always chooses the highest level of reusable work that preserves correctness.

In some situations, this completely eliminates another LLM invocation.

In others, it may reuse only the retrieval stage while generating a fresh response.

This adaptive approach minimizes latency, reduces infrastructure cost, and maintains response quality without requiring changes to an application's existing retrieval pipeline.

Design Philosophy

Remem follows a few simple principles.

Correctness before optimization

Always build the correct solution before making it faster.

---

Measure before optimizing

Every optimization should be supported by benchmarks.

---

Keep public APIs simple

Internal architecture may evolve.

The public API should remain stable.

---

Observe rather than replace

Remem should integrate with existing AI stacks instead of forcing engineers to adopt new databases or retrieval systems.

---

Current Status

Current Version:

v0.6.0-alpha

Implemented so far:

• ✅ Decoupled Execution Contexts (ExecutionContext)

• ✅ Policy-driven compatibility checks (ReusePolicy)

• ✅ Metadata filtering before similarity math (MetadataMatcher)

• ✅ Modular folder structure for reuse layers

• ✅ Policy-compliant client interface

• ✅ Metrics Collector (MetricsCollector / MetricsSnapshot)

• ✅ Persistence Layer (durable JsonStorage + ephemeral InMemoryStorage)

• ✅ Pip-installable package (pip install remem)

Not yet implemented (Planned for future versions):

• Explanations API

• HTTP server

• Language SDKs (Java, Rust)

• Distributed mode

• Rust acceleration

---

Current Architecture

The architecture cleanly decouples the public Client, policy engines, metadata matchers, and underlying storage/similarity layers, ensuring high modularity and scalability.

---

Repository Structure

remem/
├── docs/
├── remem/
│   ├── models/
│   │   ├── execution_context.py
│   │   ├── execution_record.py
│   │   └── execution_result.py
│   ├── reuse/
│   │   ├── engine.py
│   │   ├── matcher.py
│   │   └── policy.py
│   ├── similarity/
│   └── storage/
├── examples/
├── benchmarks/
└── README.md

---

Running the Examples

Clone the repository and install in editable mode (with dev extras):

git clone https://github.com/harshvardhansingh7/remem.git
cd remem
pip install -e ".[dev]"

Run the end-to-end RAG work-reuse example:

python examples/rag_reuse.py

Run the durable persistence example:

python examples/persistent_storage.py

---

Running Tests

pytest

---

Development Roadmap

v0.1.0-alpha

• RetrievalEntry
• Similarity Engine
• In-memory Storage
• Unit Tests

---

v0.2.0

Public Remem API

Remem()

↓

find_similar()

↓

Storage

↓

Similarity

---

v0.3.0

Intelligent Execution Reuse Engine with rich execution records (get_or_compute).

---

v0.4.0 (Current)

Metadata-aware policy matching engine with explicit ExecutionContext and ReusePolicy.

---

v0.5.0

Persistence Layer

Support durable storage instead of in-memory only.

---

v0.6.0

Performance Optimization

• Faster similarity search
• Profiling
• Benchmarking
• Memory optimization

---

v0.7.0

SDKs

• Python
• Java
• Rust

---

v1.0.0

Production-ready AI Work Reuse Engine

---

Long-Term Goals

Remem aims to support:

• Retrieval reuse
• Agent memory reuse
• Prompt context reuse
• SQL query reuse
• Tool execution reuse
• Knowledge versioning
• Distributed deployments
• High-performance storage engine
• Rust acceleration

---

Why Not Just Use Redis?

Redis is an excellent key-value cache.

Remem solves a different problem.

Redis answers:

"Have I seen this exact key before?"

Remem aims to answer:

"Have I already performed similar expensive AI work before?"

Instead of exact key matching, Remem focuses on semantic similarity and work reuse.

---

Learning Goals

Remem is also a personal engineering journey.

This project is being built from first principles to deeply understand:

• Distributed systems
• Storage engines
• Database internals
• Concurrency
• Memory management
• Networking
• Performance optimization
• AI infrastructure
• Open-source engineering

The goal is not simply to build another cache.

The goal is to build a useful infrastructure project while understanding every layer involved.

---

Contributing

Contributions are welcome.

As the project is still in its early stages, architecture discussions and feedback are especially valuable.

Please read CONTRIBUTING.md before opening issues or pull requests.

---

License

Licensed under the Apache License 2.0.

See the LICENSE file for details.

---

Project Status

⚠️ Early Alpha (v0.4.0)

The project is under active development.

Breaking API changes are expected until the first stable release.

---

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 interests you, consider giving it a ⭐ and following its progress.