Most 'Agent Memory' Is Just Caching With a Better Sales Deck

Letta, Mem0, Zep, Cognee, Graphiti — I have shipped with four of them, and the honest answer is that they are caching with a 10x markup and a more impressive sales deck. We have known how to remember things in software for fifty years. We just do not like the boring word for it.

What They Actually Do

Strip the SDK wrapper off any agent-memory product and you find the same five things:

• A key-value store. Postgres, Redis, sometimes a vector DB.
• A similarity lookup on read. Nearest-neighbor, HNSW, the same algorithm you shipped in your 2023 RAG project.
• A summarization pass on write. Usually a small LLM call, often batched, often over budget.
• An eviction policy. Time-decay, importance scoring, session rollup — all the words for LRU with extra steps.
• A background job to re-embed or re-summarize. Always billed.

That is the product. There is no "memory" in the cognitive-science sense. There is a write path, a read path, a refresh job, and a schema. The Letta paper reads like a cache-hierarchy paper that got dressed up for NeurIPS.

The Pricing Tells the Story

Letta Cloud starts at roughly $0.10 per million tokens. Mem0 Pro is enterprise-priced in the same neighborhood once you include embedding calls. Cognee, Zep, Graphiti — all the same shape. Compare to Upstash Redis at $0.20 per 100K requests, or pgvector on a managed Postgres for a few dollars a day. The "memory" tax is 10x to 100x what you would pay for the same primitives under their honest names. You are paying for the word "memory," the dashboard, and the SDK.

The Taxonomy Is the Con

Every agent-memory startup ships a taxonomy. Episodic. Semantic. Procedural. Reflective. Salient. This is a 1980s cache hierarchy — L1, L2, L3, backing store — wearing a 2026 cognitive-science costume. Same idea, different vocabulary. Fast tier for hot, slow tier for cold, with promotion and demotion policies. The first time I read "episodic vs semantic memory" in a vendor doc, I laughed. Cognitive-psychology 101, and the implementation is a Postgres table with a created_at column and a similarity index. There is no AI in it. There is no psychology in it. There is a schema and a SELECT statement.

The Counter-Argument I Have to Address

"Sure, but agents actually forget things, and vanilla Redis does not solve that." True. A bare KV store does not decide what to remember, summarize a 40-turn conversation coherently, or score the importance of one event over another. That is a real problem.

But it is a model problem, not a memory problem. The model has to be good at summarization and importance scoring. The right architecture is a small LLM call on each turn that decides what to write — not a vendor charging $0.10 per million tokens to wrap that call in an SDK. The agent-memory category is selling you the second half of the value at 10x the price of building it yourself with a 200-line script and a Postgres table.

The "memory" vendors want you to believe there is some magical structure — a knowledge graph, a temporal edge list, a "reflective" loop — that gives them an unfair advantage. There is not. They run a write, an embed, and a cosine lookup. You can do the same thing on Saturday morning.

The Prediction

By Q3 2027, three of the five "agent memory" unicorns will have rebranded. "Memory" will become "context graph," "agent state platform," or "long-term context layer" — anything that sounds like infrastructure and not like a cognitive feature. The products will be unchanged. The pitch decks will get longer. Customers will keep paying because nobody wants to be the engineer who said "we just use Redis and pgvector with a 200-line script" in a board meeting.

We solved memory in 1965. It is called a database. We solved caching in 1985. It is called LRU. We solved vector retrieval in 2017. It is called HNSW. The agent-memory category is a multi-billion-dollar rebrand of those three ideas.

Build the script. Skip the platform. Your agent will not remember more, but your finance team will remember the savings.