Skip to content

docs(exploration): prior-art survey for L2 runtimes + L3 cognition #24

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
cyber-ayi merged 1 commit into from
Jun 3, 2026
Merged

docs(exploration): prior-art survey for L2 runtimes + L3 cognition #24

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion ROADMAP.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,7 +83,7 @@ Each downstream task carries its own small decision-gate. None block T1.

| Open decision | Surfaces in | Status |
|---|---|---|
| harness **runtime**: OpenClaw vs ElizaOS vs Hermes | T2 / T5 | candidates listed in `harness/README.md`, unchosen |
| harness **runtime**: OpenClaw vs ElizaOS vs Hermes | T2 / T5 | unchosen; **lean = ElizaOS-core + OpenClaw-gateway + Hermes-fallback** per `exploration/prior-art-l2-l3.md` |
| **IPC transport** TS↔Python (how JSON crosses the seam) | T5 | explicitly deferred by ADR-0006 |
| Melvor **act path**: mod API vs CDP / headless-Chromium | T2 | `TASKS.md` requires choose-and-justify |
| **salient push** channel: ntfy vs Discord | T3 | **converging → MVP: AstrBot over Discord, two-way** (transport-not-brain; `discord.py` fallback). See `exploration/gateway-selection.md` |
Expand All @@ -96,6 +96,7 @@ Evaluations still in motion live in `exploration/` and resolve into ADRs.
|---|---|---|
| `exploration/substrate-selection.md` | **resolved → adr-0007** | substrate ladder: Melvor P0, Stardew P1+ (dual-control staged to P1) |
| `exploration/gateway-selection.md` | converging | operator↔agent channel; MVP = AstrBot over Discord, two-way (transport-not-brain; `discord.py` fallback) |
| `exploration/prior-art-l2-l3.md` | converging | L2 runtimes + L3 cognition survey; feeds T2 (runtime) and T4 (drive design): reuse-vs-refuse, drop the optimization objective |

## Governance / ops track (parallel to the product)

Expand Down
93 changes: 93 additions & 0 deletions exploration/prior-art-l2-l3.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,93 @@
---
topic: prior-art-l2-l3
status: converging
date: 2026-06-03
related-adrs: [adr-0004, adr-0006]
resolves-to:
---

# Prior art — L2 runtimes & L3 cognition (what to build on, what to refuse)

> Status: **converging**. A survey of current (2026) community/research work for
> the harness runtime (L2) and the cognition layer (L3), mapped to our needs.
> Feeds T2 (runtime choice) and T4 (drive-layer design). Resolves into ADRs when
> those are committed.

## Question

What existing L2 runtimes and L3 cognition systems do we stand on, and which
carry biases (reward/optimization) we must **not** adopt given autotelic-not-reward
(`adr-0004`)?

## Criteria

- **L2 harness:** runtime + perception↔action loop, local-model friendly, async
operator channel, tolerant of dual-control; **no built-in reward/curriculum bias**.
- **L3 cognition:** autotelic-not-reward, identity + needs/drives + memory/reflection,
long-horizon *non-optimal believable* behavior.

## L2 — harness / runtime / game-playing agents

| Impl | What | Fits | Don't / risk |
|---|---|---|---|
| **ElizaOS** | TS agent runtime; plugin = actions/providers/**evaluators** loop; PostgreSQL memory; Worlds/Rooms; targets game NPCs/companions | best cognition-friendly harness; TS (adr-0006); "agent-in-world" precedent | heavily **Web3/crypto**-oriented now — use only the core loop |
| **OpenClaw** | TS/Node/Electron; messaging-platform UI; **browser relay over CDP**; Ollama | browser/CDP + **operator async gateway** | heavy (Electron/sandboxes); CDP moot for SMAPI/MCP substrates |
| **Hermes** (NousResearch) | provider-agnostic, **local-first** (Ollama/LM Studio) | local "fallback brain" (privacy routing, adr-0006 rule 6) | needs ≥64k ctx; no world precedent |
| **Voyager / ODYSSEY / MindForge / Co-Voyager** | Minecraft open-world agents: **auto-curriculum + skill library** | reference architecture for open-world exploration | **skill-acquisition + curriculum = reward-shaped → conflicts with adr-0004**; reference only |
| **PIANO** (Project Sid / Altera) | parallel multi-stream cognition; agents **generate own goals** from social motivation; 1000+ agents | good **L3** reference (multi-stream + self-generated goals) | civilization/multi-agent framing; not a single-dyad harness |

## L3 — cognition / memory / autotelic motivation

| Impl | What | Fits | Gap / risk |
|---|---|---|---|
| **Generative Agents** (Park 2023) | memory stream + **reflection** + planning | memory/reflection **paradigm reusable directly** | behavior still serves schedules/goals |
| **Letta/MemGPT · Mem0** | tiered memory (core/recall/archival), auto promote/compress | **memory infrastructure off the shelf** | memory layer only; no drive/identity |
| **autotelic line** (Colas; **MAGELLAN** ICML'25; "Beyond Utility" NeurIPS'25) | agents self-generate NL goals; MAGELLAN uses **learning-progress (LP)** to guide goal choice | goal-generation machinery = the selector's theoretical core | **LP is an intrinsic reward → still optimizing**; borrow the mechanism, **drop the optimization objective** |
| **needs / personality / artificial life** (**Sophia** 2512.18202; "personality from **needs alone**"; evolving_personality; SPeCtrum) | personality/behavior **emerging from basic needs**; persistent identity | closest to the identity+needs engine | social-emergence framing; known **persona drift** + convergence to "average persona" |

## Two tensions

1. **Optimization bias is everywhere.** Open-world agents (Voyager/PIANO) and
"autotelic RL" (Colas/MAGELLAN learning-progress) ultimately **optimize**
(skills / curriculum / LP). Our stance ("purposeless, non-optimal, an end in
itself") is *more radical*. → borrow goal-generation machinery, **deliberately
discard the optimization objective**, or the drive layer collapses back into a
reward maximizer (`adr-0004`).
2. **The gap is confirmed.** "Persistent personality/needs sustaining long-horizon
*non-optimal believable* behavior" is essentially unstudied — exactly the niche
`design/autotelic-drives.md` claims. The 2026 needs/artificial-life line is the
closest but still social-emergence-framed and drift-prone. **This unfilled niche
is the project's contribution.**

## Recommendations

- **L2 (don't build a runtime — adopt one):** **ElizaOS** core (cognition-friendly
loop + memory + world precedent, TS) — *strip the crypto*; **OpenClaw** for
browser/CDP + the operator async gateway when needed; **Hermes** as the local-first
fallback brain. Voyager/PIANO are **reference architectures only** (reward bias).
- **L3 (self-build — it's the IP — but stand on giants):** reuse **Generative Agents**
memory+reflection + **Letta/Mem0** for storage; take goal-generation from
**Colas/MAGELLAN** but **cut the learning-progress reward**; take identity+needs
from the **needs-emergence/Sophia** line. Spend the budget on the unfilled niche:
long-horizon, non-optimal, believable inhabitation.

## Per-task relevance

- **T2 (Melvor adapter / runtime):** the L2 table + lean (ElizaOS-core / OpenClaw-
gateway / Hermes-fallback) is the input to the `harness runtime` decision-gate
(ROADMAP). Note: for Melvor/Stardew (mod/SMAPI/MCP adapters) OpenClaw's CDP edge
is less decisive.
- **T4 (drive layer):** the L3 table + the two tensions are the design backdrop —
*what to reuse, what to refuse*. The "drop the optimization objective" rule is the
load-bearing constraint, alongside `adr-0004`.

## Open items

- Commit the L2 runtime choice → a `harness runtime` ADR (resolves the ROADMAP gate).
- T4 will likely spawn its own exploration (selector design without an optimization
objective; memory/identity stack choice).

## Sources

- L2: [ElizaOS](https://www.elizaos.ai/) · [ElizaOS/OpenClaw/Hermes compared](https://innfactory.ai/en/blog/openclaw-vs-hermes-agent-comparison/) · [OpenClaw browser harness](https://openclawlaunch.com/guides/openclaw-browser-harness) · [Hermes Agent](https://github.com/nousresearch/hermes-agent) · [Voyager](https://voyager.minedojo.org/) · [ODYSSEY](https://openreview.net/pdf?id=vtGLtSxtqv) · [MindForge](https://arxiv.org/pdf/2411.12977) · [Project Sid / PIANO](https://arxiv.org/abs/2411.00114)
- L3: [Generative Agents](https://arxiv.org/pdf/2304.03442) · [Letta/MemGPT vs Mem0](https://vectorize.io/articles/mem0-vs-letta) · [Augmenting Autotelic Agents w/ LLMs (Colas)](https://proceedings.mlr.press/v232/colas23a/colas23a.pdf) · [Colas publications (MAGELLAN)](https://cedriccolas.com/publications/) · [LLM Agents Beyond Utility](https://arxiv.org/abs/2510.14548) · [Sophia: Persistent Agent Framework for Artificial Life](https://arxiv.org/pdf/2512.18202) · [Personality from needs alone](https://www.eurekalert.org/news-releases/1099709) · [SPeCtrum identity](https://arxiv.org/pdf/2502.08599)
Loading