Chatty DocumentationResearch modules (M0–M4, M6)
When to read this: You need a per-paper view of what is being built, where it lands in Chatty, and how far along the pipeline each module is.
Parent frame: Paper → experiment → product.
App bridge: App ↔ research bridge — maps production components (memory, context window, agent loop, …) to these modules.
flowchart TB
M0[M0 Trace contract] --> M1[M1 ReAct]
M1 --> M2[M2 AFlow]
M1 --> M3[M3 GEPA]
M1 --> M4[M4 ACE]
M2 --> M3
M0 --> M2
M0 --> M3
M0 --> M4
M1 --> M6[M6 Retrieval]
M4 -.memory store.- M6
| Module | Paper | Crate(s) | Stage | Promotion |
|---|---|---|---|---|
| M0 Trace | — (contract layer) | chatty-trace | Not started | Pending |
| M1 ReAct | Yao et al. ICLR 2023 | chatty-core, chatty-trace | Loop ships; fidelity WIP | Loop = default |
| M2 AFlow | Zhang et al. ICLR 2025 Oral | chatty-flow, chatty-optimize | Not started | Pending |
| M3 GEPA | Agrawal et al. ICLR 2026 Oral | chatty-optimize | Not started | Pending |
| M4 ACE | Zhang et al. ICLR 2026 | chatty-playbook | Not started | Pending |
| M6 Retrieval | Retrieval literature (BM25, RRF, rerankers, packing) | chatty-core tools, chatty-optimize eval | Eval built; improvements gated | Pending |
M5 (DGM) lives in the separate agenticloop repo — not documented here. See
RESERVED.md.
Status legend
| Field | Values |
|---|---|
| Stage | not started · stage-a · stage-b · promotion-pending |
| Promotion | pending · rejected · setting · default — recorded by the human reviewer only (the archived promotion-log template shows the fields) |
Cross-module work
Head-to-head comparisons (ACE vs GEPA rollout counts, combined workflows) belong in AGE-21 — human interpretation only. Agents build harness plumbing; the human reviewer runs and writes up results.