Agent Memory: 7 Types, and 2 don't actually remember
Your agent does not have a memory problem. It has seven different types of memory. Most teams only build two.
The first thing you must understand: the model remembers nothing. An LLM is a pure function. It takes an input and gives an output. It carries no state between calls. What feels like memory is just a layer re-sending history with every request. You pay for those tokens every single time.
Most engineering efforts collapse into two patterns: conversation history and RAG. These are two of the seven types. The problem? They do not make your agent smarter over time.
Here are the seven types of memory:
• Working: Everything in the current context window. • Semantic: Facts, preferences, and domain knowledge. • Episodic: A log of past events and what worked or failed. • Procedural: Skills, workflows, and tool patterns. • Retrieval: Pulling knowledge via similarity search. • Parametric: Knowledge baked into the model weights. • Prospective: Future intentions and scheduled tasks.
Two of these are not real memory. RAG is just a delivery mechanism. It is the plumbing, not the water. It moves data from a store into the working memory. If you only use a vector database, you built a pipe and forgot the liquid.
To build an agent that actually learns, you need the consolidation loop. This means turning episodic memory into semantic memory.
The process works like this:
- The agent experiences an event (Episodic).
- The agent sees the same pattern repeat many times.
- The agent abstracts that pattern into a permanent rule (Semantic).
Now, the agent does not need to reason through twelve examples. It simply applies one fact.
How to prioritize your build:
- Manage working memory as a budget. It is your highest cost. Use summarization and eviction early.
- Separate your stores. Keep facts, events, and rules in different places.
- Use a scheduler for prospective memory. Do not use a vector store for things that need to happen on a specific date.
- Draw a hard line for parametric memory. Use the model for reasoning, but use your own stores for volatile data like interest rates or product rules.
Most agents today are just a context window and a vector DB. The agents that win are the ones that can turn yesterday's mistakes into tomorrow's rules.
Source: https://dev.to/shudiptotrafder/agent-memory-7-types-and-2-of-them-arent-memory-6oi
Optional learning community: https://t.me/GyaanSetuAi
