AgentMemoryDB

Lightweight, embeddable memory backend for AI agents.

Short-term conversation buffer + long-term knowledge store + semantic search + PII masking — all in a single pip install, no server required.

---

Install

pip install agentmemodb                     # core (SQLite embedded, zero config)
pip install agentmemodb[openai]             # + OpenAI embeddings
pip install agentmemodb[huggingface]        # + HuggingFace sentence-transformers
pip install agentmemodb[langchain]          # + LangChain integration
pip install agentmemodb[langgraph]          # + LangGraph integration
pip install agentmemodb[remote]             # + connect to AgentMemoryDB server
pip install agentmemodb[all]                # everything

Quick Start

Embedded Mode (SQLite, zero config)

import agentmemodb

db = agentmemodb.Client()                       # stores in ./agentmemodb_data/
db = agentmemodb.Client(path=":memory:")        # in-memory (for tests)

# Store memories
db.upsert("user-1", "pref:lang", "User prefers Python", memory_type="semantic")
db.upsert("user-1", "pref:editor", "User uses VS Code", importance=0.9)

# Semantic search
results = db.search("user-1", "What programming language?", top_k=5)
for r in results:
    print(f"  {r.key}: {r.content}  (score={r.score:.3f})")

# CRUD
mem = db.get("user-1", "pref:lang")
all_mems = db.list("user-1", memory_type="semantic")
db.delete("user-1", "pref:lang")
n = db.count("user-1")

db.close()

Remote Mode (connect to server)

db = agentmemodb.HttpClient("http://localhost:8100", api_key="amdb_...")

# Same API as embedded!
db.upsert("user-1", "pref:lang", "User prefers Python")
results = db.search("user-1", "language?")
db.close()

---

MemoryManager — Short-term + Long-term

A unified interface that combines conversation history (short-term) with persistent knowledge (long-term):

from agentmemodb import MemoryManager

with MemoryManager("user-1") as mgr:
    # ── Short-term: conversation buffer ──
    mgr.short_term.add_system("You are a helpful assistant.")
    mgr.short_term.add_user("I'm building an API with FastAPI.")
    mgr.short_term.add_assistant("Great choice! Need help with auth?")

    # Export as OpenAI-compatible message list
    messages = mgr.short_term.to_list()
    # [{"role": "system", "content": "..."}, {"role": "user", ...}, ...]

    # ── Long-term: persistent knowledge ──
    mgr.long_term.remember("pref:framework", "User prefers FastAPI", importance=0.9)
    mgr.long_term.remember("skill:python", "Expert Python developer")

    results = mgr.long_term.recall("What framework?")

    # ── Promote conversation insights → long-term ──
    mgr.promote("insight:auth", "Use JWT with OAuth2 for FastAPI")

    # ── Combined context window for LLM ──
    ctx = mgr.get_context_window("How to set up auth?", n_messages=10, n_memories=5)
    # ctx = {"messages": [...], "relevant_memories": [...], "stats": {...}}

---

LangChain Integration

from agentmemodb import Client
from agentmemodb.integrations.langchain import (
    AgentMemoryDBChatHistory,
    AgentMemoryDBRetriever,
    AgentMemoryDBConversationMemory,
    create_memory_tool,
)

db = Client(path=":memory:")

# Chat history (BaseChatMessageHistory)
history = AgentMemoryDBChatHistory(client=db, user_id="u1", session_id="s1")
history.add_user_message("Hello!")
history.add_ai_message("Hi there!")
messages = history.messages  # [HumanMessage, AIMessage]

# Retriever (semantic search → Documents)
retriever = AgentMemoryDBRetriever(client=db, user_id="u1", top_k=5)
docs = retriever.invoke("What does the user prefer?")

# Memory tool (for agents)
tool = create_memory_tool(db, user_id="u1")
# Agent can call: tool('{"action": "store", "key": "k", "content": "v"}')
# Agent can call: tool('{"action": "recall", "query": "q"}')

# Conversation memory (history + relevant knowledge)
memory = AgentMemoryDBConversationMemory(client=db, user_id="u1", session_id="s1")
variables = memory.load_memory_variables({"input": "question"})
# {"history": "...", "relevant_context": "..."}

---

LangGraph Integration

from agentmemodb import Client
from agentmemodb.integrations.langgraph import (
    AgentMemoryDBStore,
    AgentMemoryDBSaver,
    create_memory_node,
    create_save_memory_node,
)

db = Client(path=":memory:")

# Store (long-term memory for graph nodes)
store = AgentMemoryDBStore(client=db, user_id="agent-1")
store.put("user:name", "Josh — full-stack developer")
results = store.search("Who is the user?", top_k=3)
context = store.search_as_text("user preferences")  # formatted for prompts

# Checkpoint saver (persist graph state)
saver = AgentMemoryDBSaver(client=db, user_id="system")
config = {"configurable": {"thread_id": "thread-1"}}
saver.put(config, {"messages": ["Hi!"], "step": 1}, metadata={"node": "start"})
state = saver.get(config)  # restore latest checkpoint

# Memory nodes (plug into StateGraph)
recall = create_memory_node(store, input_key="input", context_key="context")
save = create_save_memory_node(store, content_key="output")

from langgraph.graph import StateGraph, END
graph = StateGraph(...)
graph.add_node("recall", recall)
graph.add_node("generate", my_llm_node)
graph.add_node("save", save)

---

PII Masking

Built-in write-time PII detection and replacement — personal data never reaches the database:

db = agentmemodb.Client(mask_pii=True)
db.upsert("u1", "contact", "Email josh@company.com, SSN 123-45-6789")
mem = db.get("u1", "contact")
print(mem.content)  # "Email [EMAIL], SSN [SSN]"

Detects: email, phone, SSN, credit card, IP address, passport, date of birth.

---

Embedding Providers

Provider	Install	Usage
Dummy (default)	—	Client() — hash-based, no API key
OpenAI	pip install agentmemodb[openai]	Client(embedding_fn=OpenAIEmbedding(api_key="sk-..."))
HuggingFace	pip install agentmemodb[huggingface]	Custom: see below
Custom	—	Any callable matching EmbeddingFunction protocol

Custom Embedding Example

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")

class MyEmbedding:
    @property
    def dimension(self) -> int:
        return model.get_sentence_embedding_dimension()

    def __call__(self, texts: list[str]) -> list[list[float]]:
        return model.encode(texts, normalize_embeddings=True).tolist()

db = agentmemodb.Client(embedding_fn=MyEmbedding())

---

Features

Feature	Description
Embedded mode	SQLite + NumPy, zero config, pip install and go
Remote mode	Connect to AgentMemoryDB server via HTTP
Short-term memory	Conversation buffer with thread isolation, role filtering
Long-term memory	Persistent semantic knowledge with importance/confidence scores
MemoryManager	Unified short+long-term with promote() and get_context_window()
Semantic search	Cosine similarity over embeddings with configurable top-k
Memory versioning	Content-hash dedup, automatic version snapshots
PII masking	7 built-in patterns, write-time replacement, no PII stored
LangChain	ChatHistory, Retriever, Tool, ConversationMemory
LangGraph	Store, Saver (checkpoints), Memory Nodes, StateGraph helpers
Typed memories	semantic, episodic, procedural, working
Scopes	user, project, global, session
Type hints	Full py.typed support for IDE autocompletion

---

API Reference

Client / HttpClient

Method	Description
upsert(user_id, key, content, ...)	Create or update a memory
search(user_id, query, top_k=10)	Semantic search
get(user_id, key)	Get by user_id + key
get_by_id(memory_id)	Get by UUID
list(user_id, ...)	List with filters
delete(user_id, key)	Delete a memory
count(user_id)	Count memories
close()	Release resources

MemoryManager

Method	Description
.short_term	ShortTermMemory — add_user(), add_assistant(), get_messages(), to_list(), to_string()
.long_term	LongTermMemory — remember(), recall(), get(), forget(), list_all()
promote(key, content)	Move insight from conversation → long-term
get_context_window(query)	Combined messages + relevant memories for LLM
new_thread(thread_id)	Start a new conversation thread
reset()	Clear everything

---

Full Server

AgentMemoryDB also runs as a full server with PostgreSQL + pgvector, FastAPI, WebSocket, MCP protocol, React UI, and 55+ REST endpoints. See the full documentation for details.

---

License

Apache 2.0 — see LICENSE.