Skip to main content

​
What is Cognis

Cognis is a memory layer for AI agents. It lets your agents remember things — user preferences, past conversations, decisions, facts — and recall them when they matter. When a user tells your agent “I’m vegetarian” in one session, Cognis makes sure the agent knows that next time it recommends a restaurant. When a support agent resolves a billing issue, Cognis remembers the context so the next agent in the chain doesn’t ask the same questions again.

​
How It Works

  1. You send conversation messages (user/assistant pairs) to Cognis
  2. Cognis extracts facts automatically — names, preferences, decisions, context — using LLM-powered extraction with auto-categorization
  3. Your agent searches memories before responding — Cognis returns the most relevant facts using hybrid search (vector similarity + keyword matching)
  4. Memories persist across sessions — scoped by user, agent, and session so the right context reaches the right agent for the right user

​
Architecture

Cognis Architecture

​
Core Capabilities

  • Hybrid search — Matryoshka vector embeddings + BM25 keyword matching, fused with Reciprocal Rank Fusion
  • Smart extraction — LLM automatically pulls discrete facts from conversations and categorizes them (identity, preferences, work, interests, etc.)
  • Multi-tenant scoping — owner_id, agent_id, session_id — isolate memories per user, per agent, per conversation
  • Context assembly — Combine short-term conversation history with long-term memories into a single LLM-ready context

​
Quick Example

from lyzr import Cognis, CognisMessage

cog = Cognis()

cog.add(
    messages=[
        CognisMessage(role="user", content="My name is Alice. I'm vegetarian and love hiking."),
        CognisMessage(role="assistant", content="Nice to meet you, Alice!"),
    ],
    owner_id="user_alice",
)

results = cog.search(query="food preferences", owner_id="user_alice")
for r in results:
    print(r.content)
# → "Alice is vegetarian"
Cognis is available as a hosted API (pip install lyzr-adk) and an open-source library (pip install lyzr-cognis). The core architecture is the same — see the comparison. There’s also a Claude Code plugin for persistent memory across coding sessions.

​
Get Started

Quickstart

Install and run your first memory operations in 3 minutes

Configuration

API keys, init params, search weight tuning

API Reference

Complete method signatures and return types

Cookbooks

CrewAI, LangChain, LangGraph, and Agno integrations