Isn’t a long context window the same as memory?

No. A window is per-conversation and expensive; memory is durable, selective, and cheap to retrieve. Long windows also degrade in the middle, so stuffing history in is unreliable as well as costly.

Should memory be vector search or structured rows?

Both. Structured rows for profile and procedural memory where you know the key; vector search for episodic recall where you only know the situation. Using one mechanism for everything is where most implementations go wrong.

How do I know memory is actually working?

Log retrieval, not just writes. Track how many stored items are ever read back. If that number is near zero, you have a log, not a memory layer.

How long should memories live?

Profile is permanent until edited. Episodes should decay over roughly a quarter unless promoted to semantic facts by repetition. Anything never retrieved in 90 days is a pruning candidate.

Journal

AI agents

Memory Is the Moat

Six months in, the model still opens with your company name as if introducing you to yourself. That is not a model problem. Nothing was ever remembered.

Alex Cinovoj TechTide AIMay 5, 20263 min read

Key takeaways

Most "memory" features are append-only logs nothing ever reads back.
Four types matter: profile, episodic, semantic, and procedural. They need different policies.
Consolidation and forgetting are features, not maintenance chores.
A silent write failure looks exactly like a model with no personality. Check the plumbing first.

We shipped memory before we had memory. Every run wrote rows. The dashboard showed a rising count. And the output was as generic in month six as in week one, because a foreign key mismatch meant every write silently failed and every read returned nothing. The system was performing memory for an audience of one: our metrics page.

That bug was embarrassing and instructive. The symptom of broken memory is indistinguishable from the symptom of no memory: an assistant that is competent, pleasant, and has never met you.

Four kinds, four policies

Profile. Stable identity: who you are, who you serve, what you refuse to say. Always loaded, never decays, updated deliberately.
Episodic. What happened: this post ran, that claim got pushback, this customer churned. Decays with age, retrieved by similarity to the current task.
Semantic. Facts distilled from episodes: "long-form technical posts outperform for this account 3:1." Earned, not stored raw.
Procedural. How work gets done here: approval thresholds, formatting conventions, the escalation path. Retrieved by task type.

Dumping all four into one table with one retrieval strategy is the standard mistake. Profile should never compete with an episode for window space, and an episode from March should not outrank a fact learned last week.

Forgetting is the hard half

Storage is cheap, so nobody prunes, and the retrieval surface fills with near-duplicates of the same mediocre insight recorded forty times. Relevance drops. Cost rises. The model starts averaging its own past outputs, which is how systems get blander over time instead of sharper.

A weekly consolidation pass fixes it: merge near-duplicates, promote repeated episodes into a semantic fact, drop low-importance entries that were never retrieved. We run ours on a schedule and log what it removed, because a consolidation you cannot audit is a data-loss incident waiting to be discovered.

Retrieval has to be scoped, and scoped means tenant-scoped

Memory is the highest-consequence surface in a multi-tenant product. One organization’s episode leaking into another organization’s draft is not a quality bug; it is a breach with a byline. Scope every read and write by organization at the database layer, not in application code, and test the negative case explicitly.

The most dangerous thing an agent can remember is something that belongs to someone else.

Why this is a moat

Model access is a commodity. Everyone can call the same frontier model this afternoon. What is not portable is eighteen months of an organization’s decisions, rejections, wins, and vocabulary, structured well enough to retrieve. Switching cost lives there, and so does output quality.

Which means memory should be treated as a product surface, not infrastructure. Users should be able to see what the system believes about them, correct it, and delete it. That is a trust feature and, conveniently, the fastest debugging tool you will ever ship.

I go deeper on memory architecture at my site (opens in a new tab), and the retrieval stack described here is the one running in production at TechTide AI (opens in a new tab). The moat is the corpus, not the model.

“A log is not memory. Memory is a log plus a retrieval policy plus a decay function plus the discipline to forget.”

Alex Cinovoj, TechTide AI

Questions people actually ask

Isn’t a long context window the same as memory?: No. A window is per-conversation and expensive; memory is durable, selective, and cheap to retrieve. Long windows also degrade in the middle, so stuffing history in is unreliable as well as costly.
Should memory be vector search or structured rows?: Both. Structured rows for profile and procedural memory where you know the key; vector search for episodic recall where you only know the situation. Using one mechanism for everything is where most implementations go wrong.
How do I know memory is actually working?: Log retrieval, not just writes. Track how many stored items are ever read back. If that number is near zero, you have a log, not a memory layer.
How long should memories live?: Profile is permanent until edited. Episodes should decay over roughly a quarter unless promoted to semantic facts by repetition. Anything never retrieved in 90 days is a pruning candidate.

Sources & further reading

Written by

Alex Cinovoj, Founder, TechTide AI

Alex Cinovoj builds ClawLI at TechTide AI, a fleet of specialist agents that research, draft, schedule, and grade LinkedIn content with a human approving every send. He writes about what breaks when you put agents in front of a real audience.

alexcinovoj.com techtideai.io See the pre-flight gate