Ogcode

The deciding trait is that the conversation is meant to outlive the window. Most agents make you start again when the limit arrives, which means re-explaining a task you had already explained once, and this one is built specifically so that does not happen. If your work looks like long sessions rather than short errands, that is aimed squarely at you.

You are trusting it to keep the right things, which is a bet you cannot easily check. Pick it if restarts are your main pain. Pick something established if you mostly work in short bursts.

Selective recall trades one failure mode for a quieter one. Replaying everything is expensive and obviously so; recalling a subset is cheap and fails invisibly, because a requirement from turn three that did not match the current retrieval simply is not there. The agent proceeds confidently on an incomplete brief and nothing surfaces the omission.

What it does right is plan first. A planning stage in front of the edits gives you one place where the whole intent is stated, which partially offsets the thing being described above.

1. The comparison is underspecified in three ways: long session is undefined, the baseline agent is unnamed, and the measurement procedure is absent. A saving of this size is plausible for any scheme that stops replaying transcripts, which is precisely why the number needs a method before it means anything.

2. The underlying idea is sound and not novel; per-turn retrieval over conversation history is well-studied. 3. What is not addressed anywhere is the accuracy cost, and a token figure with no quality figure beside it is half an experiment.

If the efficiency result were both real and defensible, it would be a product rather than a repository, because inference cost is the largest line item every buyer in this category is trying to cut. It is published free with no commercial surface, which tells me either the author has not considered it or the result does not hold at the size a buyer would test.

Moat: an idea, and ideas in this space have a half-life of a quarter. Likely path: absorbed as a feature by an agent with users. Position: interesting, not investable, not dependable.

The pitch aims at the one number I care about, which is inference spend across sixty engineers, and then gives me no way to verify it. There is no central reporting, so a saving would show up only as an invoice that moved, with no attribution telling me why. That is not a measurement I can defend in a budget review.

No provisioning, no policy surface, no retention statement, and no pipeline role. The supplier is one person with no support commitment. Not yet, and the burden of proof sits with the claim.

The licence is permissive and the source is Go, so a fork is mine and the thing compiles into one artefact I can drop anywhere. Nothing here calls home and nothing needs a runtime beside it. That part I like without reservation.

The packaging is the tell. No documented install path means building from source is the only route, which I do not mind and most people will. No MCP client, so my servers are outside, and no local endpoint, so the inference this is so careful about economising still leaves my network entirely.