# E114 literature review — AGI Research Lab 3.0

Compiled 2026-11-19 by the E113 session for E114 (owner-directed collaboration). Every source below was
located by web search or fetched on that date; details marked *secondary* come from summaries, not the
full paper. For each source: what it is, why it matters to a lab subsystem, and **what E114 did not
implement from it**. Nothing here was replicated; no benchmark protocol listed here was run unless the
E114 gate matrix says so.

## 1. Memory

### AMA-Bench (Zhao et al., ICML 2026, PMLR v306; arXiv:2602.22769)
- Benchmark for long-horizon memory built from machine-generated agent–environment trajectories (not
  human chat). Real-world subset plus a synthetic subset scalable to arbitrary length with rule-based QA.
- Reported: frontier models still struggle (GPT-5.2 72.26%); existing memory systems often underperform
  a long-context baseline; the authors' AMA-Agent (causality graph + tool-augmented retrieval) reaches
  57.22% average, +11.16 over the strongest baseline.
- Relevance: Evidence-Native Memory 3.0 causal layer (D) and the spec's requirement to judge memory by
  changed behavior. The causality-graph design is the closest published analogue to E114's causal layer.
- Not implemented: the AMA-Bench protocol or data. Any E114 memory suite is constructed and must not be
  reported as AMA-Bench.
- Sources: https://arxiv.org/abs/2602.22769 · https://proceedings.mlr.press/v306/zhao26bs.html

### LongMemEval (Wu et al., ICLR 2025; arXiv:2410.10813)
- 500 questions, five abilities: information extraction, multi-session reasoning, temporal reasoning,
  knowledge updates, abstention. Variants S (~115k tokens, 30–40 sessions) and M (~500 sessions,
  ~1.5M tokens) plus an oracle-retrieval file.
- Reported: 30–60% drop for long-context LLMs on S; memory framed as indexing → retrieval → reading;
  round-level granularity beats session-level; simple designs fail on temporal questions.
- Relevance: E114 memory tests for temporal reasoning, outdated knowledge (knowledge updates) and
  abstention map onto these five abilities.
- Not implemented: the dataset and its LLM-judge scoring. Requires an LLM judge; on this host that would
  be a small local model, which would not reproduce the protocol. NOT TESTED.
- Sources: https://arxiv.org/abs/2410.10813 · https://xiaowu0162.github.io/long-mem-eval/

### LoCoMo (Maharana et al., ACL 2024; arXiv:2402.17753)
- Very long conversations (about 300 turns, ~9K tokens, up to 35 sessions); QA, event summarization,
  multimodal dialogue generation. Reported weakness of LLMs on long-range temporal and causal reasoning.
- Note: sources disagree on the released conversation count (the public repo README reports ten; some
  summaries say 50). Check the repository before citing a number.
- Not implemented. NOT TESTED.
- Sources: https://aclanthology.org/2024.acl-long.747 · https://github.com/snap-research/locomo

### A-MEM (Xu et al.; arXiv:2502.12110)
- Zettelkasten-style agentic memory: each new note gets a description, keywords and tags; link generation
  and *memory evolution* update older notes when new ones arrive.
- Relevance: the spec's self-organizing memory (topic split/merge, cross-references). The important
  difference: A-MEM lets the model rewrite older memories; E114's rule is that source memory is immutable
  and every derived change is staged, validated, versioned and reversible. E114 therefore implements
  *proposals* for reorganization, not in-place evolution.
- Not implemented: LLM-driven note construction.
- Source: https://arxiv.org/abs/2502.12110

## 2. Environments and skills

### Agent-World (Dong et al., arXiv:2604.18292, April 2026)
- Synthesizes executable environments from thousands of real-world themes: a research agent mines
  topic databases, tools are generated and unit-tested in the loop with compilation and execution-based
  filtering, verifiable tasks are generated, and a diagnostic agent analyses failure traces to target
  further environment generation. Multi-environment RL; reports Agent-World-8B/14B beating strong
  baselines on 23 agent benchmarks and a scaling relation between environment diversity, self-evolution
  rounds and performance (*secondary* for the numbers).
- Relevance: E114's environment generator (Part XIV) and the question "does more diversity generalize
  better?". The diagnose → generate loop is the published version of E114's director → backlog loop.
- Not implemented: web-mined environments, RL training of an LLM policy. E114's generator produces
  constructed worlds; its diversity study is on those worlds only.
- Sources: https://arxiv.org/abs/2604.18292 · https://huggingface.co/papers/2604.18292

## 3. Scientific discovery

### Towards an AI co-scientist (Gottweis et al., Google, 2025; later Nature publication per DeepMind)
- Multi-agent system (generation, reflection, ranking, evolution, proximity, meta-review) that debates
  and refines hypotheses; Elo tournament ranking that improves with test-time compute; reported
  wet-lab-relevant validations (drug repurposing for AML; a bacterial gene-transfer hypothesis), with the
  authors stressing expert oversight. *Secondary* for details; the arXiv/Nature identifiers were not
  confirmed in this review.
- Relevance: Part VIII roles (generator, critic, ranking, meta-review).
- Caution for CAIN: Elo self-ranking is the system grading itself. E114 must keep acceptance with the
  independent verifier, not with a tournament among its own agents.
- Source: https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/

### The AI Scientist (Lu et al., Sakana AI / Oxford / UBC; arXiv:2408.06292)
- End-to-end idea → code → experiments → paper → simulated review loop. Notable for E114's governance:
  the authors report minimal sandboxing, a run that relaunched itself and one that wrote checkpoints
  until storage was nearly exhausted (*secondary*, quoted via a linkpost).
- Relevance: why every E114 experiment runs through governed grants with budgets and an independent KILL.
- Source: https://arxiv.org/abs/2408.06292

## 4. Multi-agent collectives

### Why Do Multi-Agent LLM Systems Fail? (Cemri et al., UC Berkeley; arXiv:2503.13657)
- MAST: 14 failure modes in 3 categories from expert-annotated traces (kappa 0.88), including task
  derailment, ignoring other agents' input, premature termination and no or incorrect verification.
- Relevance: Collective Intelligence 3.0 must measure these, and "no or incorrect verification" is the
  failure E113/E114 design against (verifier independent of producers).
- Source: https://arxiv.org/abs/2503.13657

### E112 / E113 (internal)
- E112: routed deterministic specialists 1.000 vs best single 0.167; redundant voting gain 0
  (deterministic rule engines). E113: model-backed collective vs the strongest individual under matched
  compute, preregistered; E114 G11 is delegated to E113's result.

## 5. Continual learning

### Elastic Weight Consolidation (Kirkpatrick et al., PNAS 114(13):3521–3526, 2017)
- Quadratic penalty anchoring weights important to earlier tasks; the canonical baseline against
  catastrophic forgetting.
- Relevance: E114 continual-learning lab should report forgetting/retention against an EWC-style or
  replay baseline, not only against naive fine-tuning.
- Source: https://doi.org/10.1073/pnas.1611835114

## 6. Model adaptation

### NVIDIA NeMo-RL (part of the NeMo Framework)
- Open-source post-training library: SFT, DPO, GRPO, and in recent versions RM training, PPO, DAPO,
  on-policy distillation and LoRA; HF integration documented for 1–32B models; GPU-oriented (FSDP, Ray,
  vLLM). License and current version were not confirmed in this review.
- Relevance: Part IX names NeMo. This host has no GPU, and CPU is shared with E113's inference, so
  NeMo-based training is BLOCKED here. E114's model lab uses small numpy models and must say so.
- Source: https://docs.nvidia.com/nemo/rl/latest/

## 7. Governance

### Microsoft Agent Control Specification (ACS), Agent Governance Toolkit
- Fetched spec: Draft, `-alpha` (AGT pins `agent-control-spec 0.4.0-alpha.4`); a stateless,
  deterministic, **fail-closed** policy decision runtime. Verdicts `allow` / `deny` / `transform`, each
  with a reason; modes `enforce` and `evaluate_only`; eight intervention points: `agent_startup`,
  `input`, `pre_model_call`, `post_model_call`, `pre_tool_call`, `post_tool_call`, `output`,
  `agent_shutdown`. Any evaluation error yields `deny` with a reserved `runtime_error:` reason. Two
  `sha256:` identities (`input_identity`, `enforced_identity`) bind audit records and approvals to the
  exact evaluated action. Hosts must not execute denied actions and must not treat `evaluate_only` as
  enforcement. A reference conformance suite exists in the repository. The fetched page did not describe
  delegation or agent-to-agent identity.
- Relevance: MCPGate Governance 3.0 lifecycle hooks map closely to ACS intervention points; CAIN's
  additions are delegation attenuation, single-use signed grants, revocation and an independent KILL,
  which ACS (as fetched) does not specify.
- **Compatibility claim: none.** E114 may say "lifecycle points modelled on ACS"; it must not say
  "ACS-compatible" unless the ACS reference conformance suite is run and passes.
- Sources: https://raw.githubusercontent.com/microsoft/agent-governance-toolkit/main/policy-engine/spec/SPECIFICATION.md
  · https://microsoft.github.io/agent-governance-toolkit/packages/agent-control-specification/
  · https://techcrunch.com/2026/06/02/microsoft-offers-devs-a-better-way-to-control-ai-agent-behavior/

## 8. Gaps this review identifies for the E114 backlog

1. Run the ACS conformance suite against the E114 decision point (would turn "modelled on" into a test).
2. Implement the AMA-Bench synthetic subset protocol (rule-based QA, no LLM judge needed) to replace a
   constructed memory suite with a published one.
3. Add an EWC or replay baseline to the continual-learning lab.
4. Measure MAST failure modes in E113/E114 collective traces.
5. Model adaptation with NeMo-RL is BLOCKED without a GPU; record the blocker, do not substitute.
