the working thesis
The dominant story about why AI coding agents fail is that the models aren’t smart enough yet. Watching real harnesses do real work, we think the failures cluster somewhere else: the wrong context is in the window. We build the system that decides what the agent sees — before the prompt ever runs.
Memory for agents is not a database problem and not a model problem — it’s a context management and injection problem. What gets loaded before the prompt, what gets evicted between turns, what gets consolidated overnight, what gets quietly thrown away.
Operating systems solved the analogous problem in the 1960s with a cache hierarchy. Brains solved it on a longer timescale with hippocampal consolidation. Neither used a single flat store — and neither do we. The mechanisms we borrow from cognitive neuroscience aren’t decorations; each one ruled out a shape the system might otherwise have taken.
Coding agents forget — or remember the wrong thing. Reverie is the system that decides what an agent should remember, where to put it, and when to let it go: a placement-aware memory hierarchy, a derivability gate before every write, and an offline consolidation cycle modeled on hippocampal sleep.
A reverie mesh is a daemon, a memory store, and a shifting set of agent sessions doing real work. Cortex is the one console you drive it from — bring the mesh up, watch it live, gate what it is allowed to do. Reverie answers what an agent should know; cortex answers what the agents are doing, and what you are letting them do.
If memory were really solved, what would change about the editor? Dreamcode wraps every agent in reverie’s context — any backend inherits memory continuity — and exposes the daemon’s live activity as a first-class view. It turns out that being able to see the memory layer changes what people will trust it with.
fig. 1 — retrieval on the LoCoMo benchmark: 812 questions across five long-form conversations, against a BM25 reference on the same questions.
| metric | reverie | BM25 | Δ |
|---|---|---|---|
| R@1 | 52.8% | 46.1% | +6.7pp |
| R@5 | 77.0% | 66.5% | +10.5pp |
| R@10 | 84.0% | 71.9% | +12.1pp |
| MRR | 0.628 | 0.551 | +0.077 |
Retrieval recall, not end-to-end answer F1 — the gains concentrate where vocabulary mismatch dominates. Full breakdown on the reverie page.
The largest lifts show up on multi-hop questions (+14.9pp R@5) and adversarial ones (+11.7pp) — exactly where dense retrieval should pull ahead of pure lexical search. Where BM25 already does well, reverie still adds +7.1pp.
End-to-end F1, with judge agent, ships with the v1.0 milestone — targeted September 2026.