PaperPublished5 September 2026
Reverie: A Biological Memory Architecture for AI Agents
The second architecture, written up: conclusions as units, two brains, review, confidence, contradictions, grounding, and retrieval on request.
Research
Reverie comes out of Engineering Cognition, a research program that asks whether the understanding built during engineering work can be extracted, kept and revised by a system. We publish what we find, including when it doesn’t work.
Lineage
The story so far
- Jun 2026The question.
Working paper Engineering Cognition (30 Jun).
- Jul 2026First system, first benchmark.
Repository memory loaded into each session; EC-Bench Technical Report 001 (15 Jul) and Short Report (18 Jul). The finding that mattered: loaded memory anchored the agent.
- Jul 2026What deserves to become memory?
Notebook: Everything till now (29 Jul).
- Aug 2026Second architecture.
Conclusions as units, two brains, review, confidence, contradictions, grounding, retrieval on request. Benchmark run (13–15 Aug): the gap narrowed; the baseline is still ahead.
- Sep 2026The paper.
Reverie: A Biological Memory Architecture for AI Agents (5 Sep).
- NextWhat comes next.
The continuity experiment, a dedicated extraction model, a larger benchmark.
Featured
Start here
Evaluation
How we test it
- Benchmark
Method, results and limits.
- ExperimentThe same agent, with and without Reverie
A controlled comparison. Not yet run.
Notebook
Research log
- Notebook
A research log of the thinking behind the first version. Kept as written, with an editor’s note.
Archive
Earlier documents
Earlier documents call the first version the Engineering Memory System (EMS).
- ArchiveEngineering Cognition
The working paper that posed the question. See the paper for where the research went.
- ArchiveEngineering Memory System
Describes the first version, which loaded repository memory into each session. Superseded by the second architecture.
- ArchiveEC-Bench Technical Report 001
First benchmark run. Its aggregate gain came from the cognition-reuse metric; four of five metrics were lower with memory. See EC-Bench.
- ArchiveEC-Bench Short Report
A short report on the first benchmark run. See EC-Bench for the current picture.
Collaborate
Get in touch
Researching something related, or want to compare notes? Write to us.