cartisien-engram 0.4.0

Persistent semantic memory for AI agents — SQLite-backed, local-first, zero config

cartisien-engram

Persistent semantic memory for AI agents — Python SDK

from engram import Engram, EngramConfig

memory = Engram(EngramConfig(db_path="./memory.db"))

# Store
memory.remember("user_123", "User prefers TypeScript and dark mode", "user")

# Recall semantically — finds the right memory without exact keyword match
results = memory.recall("user_123", "what are the user's preferences?", limit=5)
# [MemoryEntry(content='User prefers TypeScript and dark mode', similarity=0.82, ...)]

---

Install

pip install cartisien-engram

Optional: LangChain integration

pip install "cartisien-engram[langchain]"

Optional: Local embeddings (recommended)

# Install Ollama: https://ollama.ai
ollama pull nomic-embed-text

Without Ollama, keyword search is used automatically.

---

Quick Start

from engram import Engram, EngramConfig

memory = Engram(EngramConfig(
    db_path="./agent.db",
    embedding_url="http://localhost:11434",  # Ollama default
))

# In your agent loop
def handle_message(session_id: str, user_input: str) -> str:
    # 1. Recall relevant context
    context = memory.recall(session_id, user_input, limit=5)
    context_str = "\n".join(f"[{e.role}]: {e.content}" for e in context)

    # 2. Build prompt + call your LLM
    response = llm.chat(f"Context:\n{context_str}\n\nUser: {user_input}")

    # 3. Store both sides
    memory.remember(session_id, user_input, "user")
    memory.remember(session_id, response, "assistant")

    return response

---

LangChain Integration

Drop-in replacement for ConversationBufferMemory:

from engram.langchain import EngramMemory
from langchain.chains import ConversationChain
from langchain.llms import OpenAI

memory = EngramMemory(
    session_id="user_abc",
    db_path="./memory.db",
    recall_limit=10
)

chain = ConversationChain(llm=OpenAI(), memory=memory)
chain.predict(input="My name is Jeff and I'm building GovScout")
# Start a new session — memory persists
chain.predict(input="What am I building?")  # Remembers "GovScout"

---

API

Engram(config?)

config = EngramConfig(
    db_path="./memory.db",       # SQLite path (default: ":memory:")
    max_context_length=4000,     # Max chars per entry
    embedding_url="http://localhost:11434",  # Ollama base URL
    embedding_model="nomic-embed-text",      # Embedding model
    semantic_search=True,        # Enable semantic search
)
memory = Engram(config)

remember(session_id, content, role="user", metadata=None)

entry = memory.remember("session_1", "User loves Thai food", "user")
# MemoryEntry(id=..., content=..., role='user', timestamp=...)

recall(session_id, query=None, limit=10, options=None)

Semantic search when Ollama available, keyword fallback otherwise.

results = memory.recall("session_1", "food preferences", limit=5)
# [MemoryEntry(..., similarity=0.84)]

history(session_id, limit=20)

Chronological conversation history.

chat = memory.history("session_1", limit=20)

forget(session_id, id=None, before=None)

memory.forget("session_1")                    # clear all
memory.forget("session_1", id="abc123")       # delete one
memory.forget("session_1", before=datetime.now())  # delete old

stats(session_id)

stats = memory.stats("session_1")
# {"total": 42, "by_role": {"user": 21, "assistant": 21}, "with_embeddings": 42}

Context manager

with Engram(EngramConfig(db_path="./memory.db")) as memory:
    memory.remember("s1", "test", "user")

---

Part of the Cartisien Memory Suite

Package	Language	Purpose
cartisien-engram	Python	This package
@cartisien/engram	TypeScript/Node	TS SDK
@cartisien/engram-mcp	TypeScript	MCP server
@cartisien/extensa	TypeScript	Vector infrastructure (soon)
@cartisien/cogito	TypeScript	Agent identity (soon)

---

MIT © Cartisien Interactive