flmnt is the memory layer for AI coding agents — it serves the decision that's true now, never a superseded one.
They remember the decision you reversed three weeks ago — and build on it. flmnt keeps your agents working from what's true: the current decision, the reasoning behind it, and nothing that's been left behind.
Everything was green. The tests passed. The agent was confident — it just couldn't know the decision it built on had moved. Context windows overflow, sessions reset, and what your team settled quietly drifts from what your agents remember.
Agents build on whatever they remember. When a settled decision changes, the old version doesn't disappear — it sits in notes, transcripts, and summaries, looking exactly as current as the new one. More context doesn't help; it's more places for the old version to live.
Your agents pick up each session already knowing what's current, what changed, and why. The Thursday agent gets Tuesday's reversal as the truth — and the old decision as history, never as instructions.
Most agent memory extracts facts from transcripts after the fact and retrieves whatever looks similar. That gets you recall — what was said. flmnt's agents author decisions, reasons, and reversals as they work, so it can serve what's true now.
When a decision is replaced, flmnt serves the new one and keeps the old one as context — never as truth. Your agents stop contradicting the team.
Every decision links to the decisions that led to it. Your agent doesn't just know what was chosen — it knows what that choice was built on.
Mistakes are first-class memory. flmnt tracks whether your agents actually stopped repeating them — and shows you.
If they're working, keep them. They stop working at the point where a decision changes. A CLAUDE.md tells your agent what was true when you last edited it; nobody rewrites the old line, so the reversed decision stays on the page and the agent keeps building on it. flmnt stores the supersession itself — ask it a question and you get the decision that's true now, not every decision you ever wrote down.
Most free memory layers do retrieval — they find the passage that matches your query. That's genuinely useful and we score the same as them on it. What they don't do is currency: when a decision has been replaced with no textual tell, similarity search returns the old one because it still reads as relevant. That's the pack our public benchmark exists to measure, results and methodology published in the open. Start on the free tier and check the claim yourself.
It's yours. Your workspace holds the decisions, explorations and mistakes your agents write; you can export the whole event stream or erase your account and everything in it from your privacy settings. We don't train models on your data.
A one-off script, a weekend project, or a single session that fits in one context window. If your agent never has to remember a decision you changed your mind about, you don't need a memory layer — and we'd rather you didn't pay for one.
A decision gets replaced with no textual tell — nothing in the words says it's dead, only the recorded relationship does. Then we ask the agent to act. Recall reads the words, and the words look current. flmnt follows the trail.
It's one of four results in our public benchmark. Every number on this page traces to it — and you can run it yourself.
The mechanism is simple to say and hard to fake: keep the record, reason over it, and hand each step exactly what it needs.
flmnt follows your agents as they work and keeps what matters: decisions, reasoning, intent, mistakes. Sessions end; the thread doesn't.
Instead of sending everything back into the prompt, a reasoning layer weighs what still holds, what's been replaced, and what this step actually needs.
Your agent gets a tight, current slice of the truth — about 4% of what's available, and 100% of what matters. Less noise, fewer tokens, grounded answers.
Retrieval degrades with scale — more history, more near-misses, more noise. flmnt inverts it: the more your team decides, the richer the trail its reasoning works from. The thread you build today makes every future answer better.
Your project's history stops being something you pay to re-read and becomes the thing your agents reason from.
Flat cost. Rising precision. And a record you can show.
Because flmnt reasons before it delivers, cost per query stays flat as your project grows — while approaches that resend history climb with every doubling. The tokens flmnt doesn't send are tokens you don't pay for, and every one is counted on a live ledger inside your workspace.
On each step, flmnt weighs everything available and forwards only the slice that bears on the task. The other 98% stays on the shelf — unsent, and unbilled.
See which decisions anchor your project, what changed and why, and exactly what flmnt saved you. Every event is immutable and replayable — a record you can show, not just describe.
If it can call a tool, flmnt can sit behind it — no SDK, no rewrite, no vendor's roadmap deciding yours. One shared memory, every agent current.
Register and tell us the thing your agents keep getting wrong — the decision that won't die, the mistake that keeps coming back. We read every one, and we'll email you when your spot opens.