Why memory is not enough

Ask anyone building agents what “memory” means and you’ll get the same answer: a place to store facts and decisions so the agent can pull them back later. A vector DB, a knowledge graph, a folder of notes, take your pick. It remembers what you discussed last week, and there are good tools that do this well now.

Now ask the harder question. What happens when the session you’re in right now crosses 180K tokens and the agent starts forgetting how it began?

The usual answer is a list of chores. Spin up a background agent. Write the plan to a file, then a second, then a third. Open a trail of GitHub issues. Leave notes in scratch markdown. Keep a tidy AGENTS.md and prompt more carefully while you’re at it. Every one of these is the same move: manually push state out of the window and hope it finds its way back. That’s you doing the filing. Maybe your tools file it for you now, straight from the conversation, so you never lift a finger. Better, but capture was never the hard part. The question is what stays in the window this turn, and saving something to a store, by hand or automatically, doesn’t decide that.

When the filing isn’t enough and the window fills mid-task, the one automatic mechanism every major agent ships kicks in: compaction. The client crushes the older turns into a lossy blob, drops the originals, and hands you back an agent that was a genius a minute ago and now can’t quite remember its own name. We all quietly accepted this for something that happens every session, often more than once: the most predictable failure in agentic coding, answered by a guillotine. So why has nobody built the thing that keeps up?

Two problems hiding behind one word

What you say	What it actually needs
”What did we decide about auth last week?”	A long-term memory store.
”Wait, what was the other thing you said we should do after this?”	Active context-window management.

These are not the same layer. A long-term store sits beside your conversation, like a notebook you keep open on the desk. It only holds what you bothered to write down, and it only helps when you reach for it. That’s exactly right for what carries across sessions: a decision, a preference, a constraint from days ago. This part has had real product attention, and it shows.

The live window is the other half, and it’s the one you’re left to manage by hand, mid-task. People do this well, but it’s a tax: every note you set down pulls your focus off the actual problem, and you have to remember to pick it back up later. And every tool for it is the same shape: static (a markdown file the model may or may not read), offline (an indexer you run between sessions), or just advice (“prompt better”). None of it is in the loop at the one moment that matters, when the window is overflowing while you work. The one mechanism that does fire on its own is compaction.

A store on the sidelines can’t intervene

That compaction step is triage with a blunt instrument: it runs whether or not it’s about to drop something you still need, with no idea what’s worth keeping. Now bolt the best long-term memory store on the planet onto the same session. What changes? Nothing. The store can answer a question if you ask it, but compaction doesn’t ask questions. It just runs. The store never gets a vote on what survives. So you can have flawless recall of last week and still watch the agent get amnesia at 200K tokens.

Some tools go further than a hand-fed store: they keep the entire conversation and let you search over it. That’s genuinely better, and it’s worth saying so. But searching is something you have to do, after you’ve already noticed something went missing, and whatever you pull back lands in the same window that was overflowing in the first place. You found the needle, and the haystack is still on fire. And automating the search doesn’t save it: async, agent-native RAG still drops what it finds into the same window, and still never decides what leaves it.

There’s a deeper reason the window wins. A model uses what’s already in front of it far more reliably than what it has to fetch: hand it the relevant text in context and it beats RAG on the same text (Li et al.), and what it does hold, it reads best at the edges, not buried in the middle (Lost in the Middle).

RAG has one honest advantage: it’s cheaper to pull a few snippets than to carry the whole history. But that only pays off when a small slice can substitute for everything else, and an agent session is the opposite case. RAG can only add to the window, never stand in for it.

What managing the window actually looks like

Two hundred turns deep, that means keeping the auth decision and the path of the file you’re editing while dropping the stale 4,000-token test dump, without anyone having to ask. To make that call on every turn, instead of letting a blunt summarizer maul the whole thing at the boundary, a few things have to be true:

• Keep a distilled prefix at the front. Compress the early conversation into a dense, structured record of what was established: the decisions, the shape of the work, the constraints. That record stays pinned to the front of the window where the model can always see it, carrying the load-bearing facts the rest of the work leans on.
• Compress gradually, in layers. Full passthrough while there’s room. As pressure builds, compress the raw turns behind the distilled prefix. Under more pressure, strip what ages worst first: stale tool output, redundant dumps. Emergency compression stays the last resort.
• Calibrate from the real token counts. The window is a hard budget. What gets cut should follow the actual token counts the API reports back, and what each model’s real context and pricing are. A character cap hardcoded once and forgotten tracks none of that.
• Protect the live edge. The most recent turns stay intact, always. Whatever happens upstream, the active end of the conversation is where the work is happening, so it stays whole.

What matters here is the loop: something is actively deciding what stays in the window on every turn, instead of a one-shot summarizer flattening the session the moment it overflows.

Two layers, one stack

This isn’t memory versus context management. You need both, and they stack:

1. Temporal storage. Everything that’s said, captured and indexed, so nothing is truly lost even after it leaves the window.
2. Distillation. That history compressed into a dense prefix, so the established facts survive context pressure and stay in front of the model.
3. Long-term knowledge. The durable decisions, patterns, and preferences pulled out and carried across sessions, retrievable on demand.

That store, layer 3, is what people mean today when they say “memory”. It’s the top of the stack, not the whole stack. Without the two layers under it, it’s a filing cabinet next to a conversation that’s quietly falling apart.

The work this is built on

We didn’t invent this architecture, and we want to be clear about that. Two teams got here first and proved it works. Sanity’s Nuum gave us the framing we still use, “distillation, not summarization”, and showed a three-tier memory staying coherent across thousands of messages. (Simen Svale’s description of compaction as “JPEG compression of memory management” is hard to forget once you’ve read it.) Mastra’s Observational Memory worked out the observer/reflector loop and the move from rigid JSON to plain, timestamped observation logs. Lore is built on both, and it’s better for it. Their write-ups are worth your time.

There’s a catch, though. Both teams shipped this architecture inside their own agent. Nuum is a standalone REPL (Anthropic-only, yolo-mode, built to power Sanity’s Miriad). Observational Memory lives inside the Mastra framework. The memory is excellent. The packaging is the constraint. Almost no one is going to replace the coding agent they already rely on just to get a context-management layer.

Unhook it from the harness

So that’s the gap: take the same ideas and unhook them from the harness. Deliver them as a layer that works with whatever agent you’re already using, on whatever provider you’re already paying for.

Here’s the distinction that actually matters, and it’s easy to miss now that everything calls itself memory. A store is a tool the agent reaches for: it sits off to the side and answers when it’s asked, however clever the asking has gotten. A layer sits in the request path. Every turn flows through it, and it can reshape that turn before the model ever sees it. That position is the only one from which you can manage the live window at all.

It’s also the most invasive seat in the system, and that’s worth being honest about. A thing that sees every token of every session only earns that seat if it’s yours: a single file on your own disk, no account to create, no database to run, nothing leaving your machine unless you choose to share it. It’s also why this is Fair Source (FSL-1.1-Apache-2.0): the code that touches your tokens is right there to read, and it turns into Apache 2.0 on a timer. That last part is hard for a cloud service to follow you into. Routing every token through someone else’s box is a lot to ask, and a lot for that box to be responsible for. The seat really only makes sense when you’re the one sitting in it.

That same vantage point, seeing every request as it happens, is also the right home for a lot more than distillation:

• Warm the cache automatically, so you’re not paying full price for a cold context at the start of every session.
• Extract long-term knowledge on its own: decisions, conventions, gotchas, without you remembering to save anything.
• Recognize patterns in how you and your agent actually work, and act on them.
• Track token spend day to day, across every agent you run.

That last one is quietly becoming the point. As model and token costs climb, the layer that sees every request is exactly where cost control belongs, and because it runs on your own machine, there’s no infrastructure bill and no per-request meter to feed. A store on the sidelines can’t do any of this. Plenty of them run the other way: a server you’re renting, an extra model call to extract and index every exchange, and a metered fee on top of the requests you were already paying for.

And once memory lives in that layer instead of inside one agent, it stops belonging to any single agent. Your context follows you. Move from Claude Code to Codex to OpenCode, or reach for a different agent on a different task, and the same memory comes along. A team running a mix shares one knowledge base instead of three separate silos. Portability comes free once the memory no longer lives inside the harness.

Memory is the wedge. The layer is the platform.

The one question worth asking

Next time something is pitched to you as agent “memory”, ask one question:

What happens at 200K tokens?

If the answer is some version of “you can search what we stored”, or worse, “write a better context file”, then it’s a storage tool, and the part that actually breaks mid-session is still yours to handle. That’s fine, and a good store is genuinely worth having. But it has no answer for the session that’s overflowing right now, which is the problem you’ll hit today, not next week.

Memory is table stakes. Managing the window, in the loop, while it fills, is the part nobody wants to build. Make sure you know which one you’re being sold.