RECA is a flexible, lightweight memory and retrieval system designed to empower AI agents with real-time recall, long-term memory, and contextual search. Built for developers building autonomous tools, chat interfaces, personal assistants, or multi-agent systems - RECA enables your agents to remember what matters and find what they need, when they need it.
Memory is not storage. It’s context, grounding, and cognition.
RECA helps agents:
- Remember past messages, actions, or plans
- Search over documents, web content, or APIs
- Store knowledge as structured objects
- Retrieve relevant info via semantic or keyword matching
- Build and update memory on the fly
- Inject memory into LLM prompts dynamically
RECA powers recall-aware agents. Instead of generating in isolation, they generate with purpose.
RECA requires Python 3.10+ and uses Poetry for dependency management.
# Clone the repo
cd reca
poetry install
Or install it directly into your Poetry project:
poetry add reca
The basic unit in RECA is a MemoryObject. Each object contains:
text: the contenttimestamp: creation or modification timesource: where it came from (file, agent, API, etc)tags: optional labelsembedding: optional vector for semantic search
from reca.memory.schema import MemoryObject
from datetime import datetime
obj = MemoryObject(
text="Bananas are high in potassium.",
timestamp=datetime.utcnow(),
source="notes.txt",
tags=["nutrition"]
)
from reca.memory.core import Memory
mem = Memory()
mem.insert("Elephants have excellent memory.")
results = mem.search("Which animals have strong memory?")
print(results[0].text)
Mix local memory with external retrievers:
from reca.retrievers.hybrid import HybridRetriever
from reca.retrievers.local import LocalRetriever
from reca.retrievers.wikipedia import WikipediaRetriever
hybrid = HybridRetriever([
LocalRetriever(mem),
WikipediaRetriever()
])
answers = hybrid.search("What is quantum entanglement?")
Split PDFs or markdown into memory-ready chunks:
from reca.parsers.markdown import parse_markdown
chunks = parse_markdown("docs/overview.md")
for c in chunks:
mem.insert(c)
Supported formats include .md, .txt, .pdf, and .json.
RECA supports pluggable embedding backends. Use OpenAI or local transformers:
from reca.models.embedding import OpenAIEmbedder
mem = Memory(embedder=OpenAIEmbedder())
You can also add your own model:
from reca.models.embedding import CustomEmbedder
class MyEmbedder(CustomEmbedder):
def embed(self, text):
return my_custom_embedding(text)
Search web content on-the-fly:
from reca.retrievers.web import WebRetriever
web = WebRetriever()
results = web.search("Latest news on GPT-5")
RECA includes pluggable retrievers for:
- Wikipedia
- Arxiv
- DuckDuckGo
- Local browser history
- Custom endpoints
Run tests locally using Poetry:
poetry run pytest
To add a custom retriever:
- Subclass
BaseRetriever - Implement
.search(query) - Return a list of
MemoryObject
To add a memory backend:
- Subclass
MemoryBase - Implement
.insert(),.search(), and.delete()
- Conversational agents with memory
- Retrieval-augmented generation (RAG)
- Autonomy stacks and planners
- Personal assistants that retain facts
- AI tools that need context over time
Interested in extending RECA? PRs are welcome!
You can contribute:
- New retrievers
- New backends (e.g., Redis, Qdrant)
- Parser support for new formats
- Improvements to documentation
- Agent integrations and demos
RECA is released under the MIT License. See LICENSE for details.
Developed independently. No affiliation with upstream projects or repositories.