AI and LLMAgentsPerformance and Scale plans

Autonomous Agents with Memory

Give long-running agents conversation context, recalled knowledge, task state and an audit trail, all in one place.

kinds of memory: 4
store for all of them: 1
agent framework: Any

Step 1 of 6

Open a conversation

A session keeps a rolling window of the most recent turns.

Example sessionplaying

# Open a conversation

hotkv> AMEM.CONV.CREATE session:42 WINDOW 20

An animated walkthrough with real HotKV commands and example replies. Click any step to jump to it.

The problem

An agent that forgets between steps cannot finish long tasks. The usual answer is a chat history table, a vector store for facts and a scratch space for state, and still no single record of what the agent did.

How HotKV handles it

HotKV provides four kinds of agent memory as commands. Conversation windows hold dialogue context, semantic memory recalls relevant facts by meaning, working memory tracks the state of the current task, and an episodic log records what the agent did.

Step by step

What happens, command by command

01
Open a conversation
A session keeps a rolling window of the most recent turns.
AMEM.CONV.CREATE session:42 WINDOW 20
02
Record what the customer said
Each turn is appended with its role.
AMEM.CONV.APPEND session:42 user "Why was my invoice declined?"
03
Recall relevant facts
Facts stored earlier with AMEM.SEM.STORE, in semantic memory made with AMEM.SEM.CREATE, are found by meaning, not by keyword.
AMEM.SEM.RECALL billing-agent 3 0.02 0.7 ...
04
Track the task
Working memory holds where the agent is, and expires when the task is long over.
AMEM.WORK.SET session:42 step "checking payment gateway" TTL 600
05
Log the action
Every action the agent takes is recorded for audit and debugging.
AMEM.EPIS.LOG billing-agent action "Reissued invoice INV-9921"
06
Context for the next turn
The recent turns go into the next prompt, so the agent never loses the thread.
AMEM.CONV.WINDOW session:42 10

Where it fits

Real-world scenarios

Customer service agents

An agent handling refunds remembers the customer's earlier messages, the store's policies and which step of the refund it has reached.

Research assistants

Findings are stored as semantic facts during a long investigation and recalled when writing the summary.

Operations bots

Every action an on-call bot takes is in the episodic log, so humans can review exactly what it did and why.

Built in

What you get

Conversation memory

A rolling window of recent turns for each session.

AMEM.CONV.APPENDAMEM.CONV.WINDOW

Semantic memory

Facts recalled by meaning, optionally with a TTL.

AMEM.SEM.STOREAMEM.SEM.RECALL

Working memory

Task state that can expire on its own.

AMEM.WORK.SETAMEM.WORK.GET

Episodic log

Actions and observations, queryable by type and time.

AMEM.EPIS.LOGAMEM.EPIS.QUERY

Reference:AMEM.CONV.APPEND AMEM.SEM.RECALL AMEM.WORK.SET AMEM.EPIS.LOG

Common questions

How is this different from storing chat history in a database?

The four memory types share one store and one connection, so recall, task state and the action log do not each need their own service.

Does the agent framework matter?

No. Memory is exposed as commands over the Redis protocol, and the MCP server exposes it to MCP clients. For LangGraph there is a checkpointer and store.

Which plans include it?

Agent memory is part of the Performance and Scale plans.