Gemento

On a 9-task cost-aware benchmark, 8-loop ABC orchestration with Gemma 4 E4B scored 78.1% versus 41.3% for the same model in 1-loop solo, and 59.1% for a one-call Gemini 2.5 Flash baseline — at the cost of roughly 20× wall time and zero per-trial API cost. Across 13 sequentially numbered hypotheses (H1–H13, 540+ trials), a recent role-axis ablation suggests a position-effect pattern: pre-stage role addition (Extractor, H11) shows Δ=+0.05 (Cohen's d=+0.32, p=0.198), while post-stage role addition (Reducer, H12) shows Δ=−0.05 (d=−0.32, p=0.180) — a directional mirroring at similar effect sizes. Both effects are not statistically significant at n=15 paired tasks; cross-model replication is planned to test whether this direction generalizes.

📚 Korean version: README.ko.md · 📄 Paper draft: docs/paper/draft.md · arXiv: TBD

Last updated: 2026-05-05

Gemento is an experimental harness, not a new architecture or a paper. It tests whether small local LLMs can run long workflows when memory, tools, roles, and control are externalized.

Why I built this

I started this while building secall and tunaFlow. The recurring problems were long-term memory, context over-spend, and work that had to survive across sessions. My first thought was simple: if I clear the prompt context and rely on database retrieval, can the database behave like near-infinite context? That line of thinking brought me back to Memento and Leonard's tattoos, polaroids, and phone calls. This repository is the side-track that grew out of that idea: not a new architecture, not a paper, but a measured notebook about whether small models improve when memory, action, critique, and control are externalized.

What this is / is not

This is:

• a reproducible experiment harness for small local LLM workflows
• a measured notebook about externalized state, tools, roles, and control
• a public baseline for reproduction and disagreement

This is not:

• a new model architecture
• a training method
• a claim that 4B models replace frontier models
• a claim that ABC+Tattoo universally beats RAG

Scope and reproducibility caveats (read before quoting numbers)

• All headline numbers come from a single base model (Gemma 4 E4B) on a small custom benchmark (15 tasks main + 10 long-context tasks). Cross-model replication is planned, not completed.
• Several Stage 5 hypotheses (H4, H10, H11, H12) are not statistically significant at n=15 paired tasks — they are reported with effect sizes (Cohen's d) and bootstrap CIs as replication targets, not as confirmed effects.
• The deterministic keyword scorer (score_answer_v3) cannot distinguish "real quality difference" from "answer-style mismatch with the keyword set". This is most relevant to H12 (Reducer compresses multi-source answers); LLM-as-judge replication is planned.
• The Exp10 vs Gemini Flash comparison is benchmark-specific and trades roughly 20× wall time for $0 per-trial API cost — not a general superiority claim.
• The "position-effect" observation from H11+H12 is a directional pattern at mirrored effect sizes, not a confirmed asymmetry.

Core idea

Gemento treats four axes of LLM cognition as externalizable — moved out of model weights and into the workflow:

Axis	What goes outside the model	Mechanism	Anchor experiment
Tattoo	Working memory (claims, evidence, status)	Structured JSON state passed across loops	Exp02, Exp09
Tools	Computation (math, search, retrieval)	OpenAI-compatible function calls (calculator, linalg, linprog)	Exp08, Exp08b
Role	Self-validation	A (Proposer) / B (Critic) / C (Judge) — separate prompts, same base model	Exp03, Exp035, Exp06
Orchestrator	Termination · phase transition · resource budget	Deterministic Python loop, not the model itself	Exp02, Exp07

This is not about replacing a 70B model with a 4B one. It is about asking what structure can extract from a small model (Gemma 4 E4B, effective 4B parameters) that single-pass inference does not surface.

What I measured

This repository tracks sequential hypothesis IDs H1 to H9c across the four externalization axes.

Note: H1-H9 below denote sequentially numbered hypotheses about externalization axes. They are not the statistical H₀/H₁ pair.

ID	Hypothesis (Axis)	Verdict	Evidence
H1	[Orchestrator externalization] Multi-step loops outperform single-pass reasoning	✅ Supported (+44.4pp Exp02; +37pp Exp10)	Exp02 v2, Exp10
H2	[Role externalization necessity, falsification] Self-validation can detect its own errors	❌ Rejected (0/15 detected)	Exp03
H3	[Role externalization] Cross-validation with separated roles can detect errors	✅ Supported (12/15, 80%)	Exp035
H4	[Role externalization synergy] A-B-C role separation outperforms repeated single-agent iteration	⚠ Conditionally supported (synthesis category only, Stage 2C 2026-05-02) — 15-task ablation: ABC > Solo-budget +0.0444 (reverses 9-task subset's Solo +0.067 direction), synthesis +0.140 (recovery driver); not significant at n=15, Cohen's d=0.449 medium. Detail: docs/reference/h4-recheck-analysis-2026-05-02.md	Exp06 + Stage 2C
H5	[Orchestrator ceiling effect] Raising MAX_CYCLES improves accuracy	⚠️ Partially rejected - the ceiling was not the limit; actual saturation appeared around actual_cycles ≈ 7	Exp07
H6	[Role externalization refinement] Phase-specialized prompts outperform the baseline	✅ Conditionally supported (+5-6pp in long loops)	Exp07
H7	[Tool externalization] External math tools compensate for E4B's calculation limits	✅ Supported (+18.3pp, math-04 0→80%)	Exp08
H8	[Tool externalization stability] Error hints and mandatory tool rules reduce tool neglect and operator confusion	✅ Supported (neglect 0%, calculator 100%, math-04 0→100%, +23.3pp)	Exp08b
H9a	[Tattoo externalization, physical limit] ABC+Tattoo (chunked) outperforms solo dump under long context	✅ Supported (+68.3pp, Solo 0% → ABC 100% at Large 20K)	Exp09
H9b	[Distinctiveness] ABC+Tattoo contributes something beyond a RAG baseline	⚠️ Inconclusive (5-trial stats NOT SIGNIFICANT p=0.798; overall Δ=+2.0pp; 3-hop only +20.0pp differentiation; Small Paradox confirmed)	Exp09
H9c	[Error mode difference] ABC fails differently from Solo and RAG	✅ Supported (Solo: format_error, RAG: wrong_synthesis, ABC: evidence_miss + wrong_synthesis)	Exp09
H10	[Role externalization — strong Judge / Mixed Intelligence] A stronger Judge C (Gemini 2.5 Flash) compensates for weaker Proposer/Critic (A/B = Gemma 4 E4B)	⚠ Inconclusive — effectively rejected (Exp11 2026-05-03). Δ(mixed − baseline) = −0.0811 (mixed underperforms baseline-all-Gemma); Cohen's d = −0.316 (small, negative); not significant at n=15, p=0.293; logic category catastrophic (−0.275). Inverse mechanism observed: stronger Judge interferes with weaker model's self-discovery chain (logic-02 case study). Detail: docs/reference/exp11-mixed-intelligence-analysis-2026-05-03.md	Exp11
H11	[Role externalization — separation / addition (Extractor Role)] A new Role (Extractor, same Gemma 4 E4B model) that pre-extracts claims/entities from the prompt and prefixes them into A→B→C input improves accuracy	⚠ Conditionally supported, power-limited (Exp12 2026-05-04). Δ(ext − baseline) = +0.0500; Cohen's d = +0.323 (small, positive); not significant at n=15, p=0.198 (Wilcoxon). Directional improvement on logic-02 (0.3→0.6) and synthesis-05 (0.55→1.0). Suggests Role-axis separation/addition may be a safer evolution path than strengthening; not yet replicated cross-model. Detail: docs/reference/exp12-extractor-role-analysis-2026-05-04.md	Exp12
H12	[Role externalization — separation / addition (Reducer Role, post-stage)] A new Role (Reducer, same Gemma 4 E4B model) that post-stage polishes the final tattoo + final_answer improves keyword-match accuracy	⚠ Inconclusive — effectively rejected, power-limited (Exp13 2026-05-05). Δ(reducer − baseline) = −0.0533 (bug-excluded) / −0.0711 (with bug); Cohen's d = −0.323 (mirror of Exp12's +0.323 in magnitude); not significant at n=15, p=0.180 (Wilcoxon). Proposed mechanism — abstraction loss: in synthesis-04, baseline produced structured multi-source analyses scoring high under the keyword scorer, while reducer produced compressed single-estimate answers scoring lower. The current keyword-based scorer (score_answer_v3) cannot distinguish "actual quality drop" from "scorer-style mismatch"; LLM-as-judge replication is planned (P1-3 in docs/reference/paper-review-action-items-2026-05-05.md). Together with H11, this provides evidence consistent with a position effect at similar effect-size magnitudes in opposite directions; cross-model replication is required before any stronger claim. Detail: docs/reference/exp13-reducer-role-analysis-2026-05-05.md	Exp13
H13	[Tool externalization — agent-active retrieval] ABC agents that actively call a search_chunks(query, top_k) BM25 retrieval tool during cycles outperform a sufficient-context baseline (32K-context window with the full document in the prompt) on long-context tasks	⚠ Inconclusive — effectively rejected, statistically significant negative direction at this scale (Exp14 2026-05-05). Δ(search − baseline) = −0.220; Cohen's d = −1.000 (large); Wilcoxon p=0.0312 / paired t p=0.0115 — first statistically significant verdict in Stage 5. Bootstrap 95% CI Δ: [−0.36, −0.10] (does not include 0). Mechanism = insufficient retrieval iterations on multi-hop tasks: a 5-trial diagnostic on longctx-large-2hop-01 showed total_tool_calls per trial = [1, 2, 2, 3, 0] → accuracy = [0, 1, 1, 1, 0]. The single-call trial recovered the hop-1 entity then prematurely concluded "the document does not contain a [number]" instead of issuing a hop-2 query. Needle (1-hop) tasks succeeded 5/5 with one well-formed call. Tool axis sub-distinction: deterministic single-call computation (calculator/linprog, H7/H8 +18~23pp) ≠ probabilistic agent-iterative retrieval (H13 −22pp). The result applies specifically to agent-active BM25 retrieval against a 32K-context baseline; a weaker form ("retrieval helps when context is the limit") is not addressed because the baseline saturated. Detail: docs/reference/exp14-search-tool-analysis-2026-05-05.md	Exp14

What worked / What didn't

• Exp02 v2: 50% to 94.4% after forcing 8 loops on the same E4B model.
• Exp03: self-validation detected 0 out of 15 planted errors.
• Exp035: role-separated cross-validation recovered to 80%.
• Exp08b: math tool calls plus error-message hints moved math-04 from 0% to 100%.
• Exp09: on Large 20K long-context tasks, Solo dump scored 0%, RAG scored 67%, and ABC+Tattoo scored 100%.
• Exp10: on a 9-task / 540-trial cost-aware comparison, the same Gemma 4 E4B went from 41.3% (1-loop) to 78.1% (8-loop ABC). On this benchmark the ABC condition outperformed a one-call Gemini 2.5 Flash baseline (59.1%) by 19 percentage points, trading API cost for roughly 20× longer wall time. The result is benchmark-specific (9 tasks, math/logic/synthesis/planning) and 1-call vs 8-loop is an asymmetric comparison; it is not a general claim that small local models replace frontier APIs. ABC infrastructure had 4 trial-level JSON parse fails (early-stop pattern, see docs/reference/exp10-v3-abc-json-fail-diagnosis.md).
• Stage 2C (2026-05-02) re-evaluated H4 with an expanded 15-task ablation. ABC outperformed Solo-budget by +0.044 (reversing the 9-task subset's Solo +0.067 direction). The synthesis category was the recovery driver (+0.140); statistically not significant at n=15 but Cohen's d=0.449 (medium). H4 verdict moved from Inconclusive to ⚠ Conditionally supported (synthesis only). See docs/reference/h4-recheck-analysis-2026-05-02.md.
• Exp11 (2026-05-03) tested Mixed Intelligence (Judge C = Gemini 2.5 Flash, A/B = Gemma 4 E4B) and the result was negative: Δ(mixed − baseline) = −0.081, Cohen's d = −0.316 (small, negative). The logic-02 case study (Δ = −0.900) was decisive — baseline produced 4/5 correct answers explicitly stating "105 inconsistent" via the inclusion-exclusion principle, while mixed produced 5/5 nulls or keyword-missing answers. Inverse mechanism observed: a stronger Judge interferes with the weaker model's self-discovery chain via Tattoo-schema mismatch and premature convergence. H10 ⚠ Inconclusive — effectively rejected. The framework's next direction shifts from Role strengthening to Role separation/addition.
• Exp12 (2026-05-04) tested Extractor Role (a new Role using the same Gemma model, pre-extracting claims/entities and prefixing them into A's input). Δ = +0.050 (positive), Cohen's d = +0.323 (small). Directional improvements on logic-02 (0.3→0.6) and synthesis-05 (0.55→1.0). Proposed mechanism: cycle-1 input organization (assisting A, not replacing). H11 ⚠ Conditionally supported, not statistically significant at n=15 (p=0.198).
• Exp13 (2026-05-05) tested Reducer Role (a structurally symmetric counterpart — same Gemma model, post-stage polishing of the final tattoo + final_answer). Δ = −0.053 (bug-excluded), Cohen's d = −0.323 — opposite in sign to Exp12 with nearly identical magnitude. In the synthesis-04 illustrative trial, baseline produced structured multi-source analyses ("## Comprehensive Analysis ... Identification of Contradictions ...") scoring 1.0, while the reducer-polished output compressed to single-point estimates ("The best estimate is 270 individuals.") scoring 0.33–0.67. Caveat — keyword scorer artifact possibility: this drop may reflect real abstraction loss OR style mismatch between compressed answers and the deterministic keyword scorer (score_answer_v3). The two explanations are not separable from the current data; LLM-as-judge replication is planned. H12 ⚠ Inconclusive, effectively rejected (not significant at n=15, p=0.180). Together with H11, the two results provide evidence consistent with a position effect at similar effect-size magnitudes in opposite directions — but the pattern is a replication target, not a confirmed asymmetry.
• Exp14 (2026-05-05) tested an agent-active BM25 search tool against a sufficient-context baseline (32K-context window holds even Large 20K-word documents) on the long-context taskset (10 tasks). Δ = −0.220, Cohen's d = −1.000 (large), Wilcoxon p=0.0312, paired t p=0.0115 — the first statistically significant verdict in Stage 5. Bootstrap 95% CI Δ: [−0.36, −0.10]. The mechanism is insufficient retrieval iterations on multi-hop tasks: in the diagnostic 5-trial run on longctx-large-2hop-01, tool-call counts of [1, 2, 2, 3, 0] mapped onto accuracy [0, 1, 1, 1, 0]. The single-call trial recovered the hop-1 entity ("Westbrook Institute") then prematurely concluded the document lacked the answer instead of issuing a hop-2 query. Needle (1-hop) tasks succeeded 5/5 with one well-formed call. H13 ⚠ Inconclusive, effectively rejected. → Tool-axis sub-distinction: deterministic single-call computation tools (calculator, linprog — H7/H8 +18~23pp on math tasks) ≠ probabilistic agent-iterative retrieval (H13 −22pp on multi-hop). Combined with the H11+H12 position effect, Stage 5's integrated finding is "more structure is not monotonically better; what matters is where it is placed and how it iterates" — the proposed paper's main claim. Stage 6 next: cross-model replication (Llama 3.1 8B, Llama 3.3 70B, Qwen 2.5 7B Q4_K_M) plus LLM-as-judge auxiliary evaluation to address the H12/H13 keyword-scorer caveats.

What clearly did not work was expecting the same model to notice its own failure mode without a role change. The most useful result in this repository is probably negative: the model does not reliably criticize itself, even when it can criticize the same reasoning from another role.

Why this matters

If you are working on externalization for small models, structured state, or multi-role validation, this repo is meant to be a public coordinate rather than a closed notebook. It is MIT-licensed on purpose. I would rather have competing measurements, forks, and disagreements in public than keep this siloed until it looks polished.

Quickstart

Environment

Tested with Python 3.14. Earlier 3.12+ should work but is not regularly verified — if you hit a syntax/typing error on an older interpreter, please report the version.

git clone https://github.com/hang-in/gemento.git
cd gemento
python3.14 -m venv .venv
source .venv/bin/activate   # Windows PowerShell: .venv\Scripts\Activate.ps1
pip install -r requirements.txt

Inference server (LM Studio or llama.cpp — both supported)

Gemento expects an OpenAI-compatible chat-completions endpoint that supports tool_calls. Two reference setups:

• LM Studio (current author's setup): load Gemma 4 E4B Q8_0 with at least 32K context window enabled, server on http://192.168.1.179:1234 (or http://localhost:1234). Set API_BASE_URL accordingly in experiments/config.py.
• llama.cpp server: ./server -m gemma-4-e4b-q8_0.gguf -ngl <max layers> -c 32768 --port 8080. Default http://localhost:8080.

Confirm the active model id and use it for MODEL_NAME:

curl -s $API_BASE_URL/v1/models | jq -r '.data[].id'

Expected: a model id such as gemma4-e4b, google/gemma-3n-e4b-it, or whatever your server reports. Set MODEL_NAME in experiments/config.py to exactly this string — a mismatch causes silent fallback to defaults on some servers.

Smoke test

cd experiments
python tools/smoke_test.py

Expected: SMOKE TEST PASSED: math-04 answer=..., tool_calls=...

Reproduce the headline Exp10 result (~12-14h on a single 8GB GPU)

The README hero number (Gemma 4 E4B 41.3% → 78.1% across 9 tasks × 60 trials = 540 trials per condition) comes from Exp10. To reproduce on your machine:

# from the gemento root, after smoke test passes
python -m experiments.exp10_reproducibility_cost.run --conditions gemma_1loop --trials 60
python -m experiments.exp10_reproducibility_cost.run --conditions gemma_8loop --trials 60
# results land in experiments/exp10_reproducibility_cost/results/
# scorer: score_answer_v3 (negative_patterns) — applied automatically

Per-trial wall time depends on hardware; on a 3060 Ti 8GB with LM Studio the 8-loop condition averages ~7-8 minutes/trial. Optional: gemini_flash_1call condition requires GEMINI_API_KEY in .env and incurs API cost (≈$0.05 for 60 trials at 2026-05 pricing).

Other experiments

python run_experiment.py baseline
python run_experiment.py tool-use
python measure.py "results/exp08_*.json" --markdown --output results/exp08_report.md

All experiments support checkpoint resume. If interrupted, rerunning continues from partial_*.json.

Research notes

If you want more than the README:

• README.ko.md — Korean full README with the same hypothesis table, plus extra sections on open questions and Korean-language scoring caveats.
• docs/reference/researchNotebook.md — Per-experiment notebook in 5W1H format (who / when / where / what / why / how). Korean.
• docs/reference/researchNotebook.en.md — English mirror of the above (kept in sync; Closed-append-only policy).
• docs/reference/conceptFramework.md — The four-axis externalization framework, including unverified candidate axes (Extractor, Reducer, Search Tool, Graph Tool, Evidence Tool, Critic Tool, Large Model Tool / Mixed Intelligence — Exp11 effectively rejected).
• docs/reference/namingConventions.md — Notation and terminology rules (Stage NX, task-NN, condition slugs, etc.).
• docs/reference/scoringHistory.md — Scorer evolution (v0/v2/v3).
• docs/reference/failureLabels.md — FailureLabel enum and failure classification standard.
• docs/reference/resultJsonSchema.md — Result JSON top-level meta v1.0.
• Analysis reports — Stage 2C H4 recheck · Exp11 Mixed Intelligence · Exp12 Extractor Role · Exp13 Reducer Role · Exp14 Search Tool · Exp09 5-trial drop.

Plan-level history (decisions and revisions) lives under docs/plans/ — see docs/plans/index.md for the running list (Active + Recently Done).

Reproduce / Extend

(a) Run with another model

Change these lines in experiments/config.py:

MODEL_NAME = "qwen2.5-7b-instruct"  # or phi-4, llama-3.2-3b, ...
API_BASE_URL = "http://localhost:8080"

The llama.cpp server must support tool_calls. Check /props for chat_template_caps.supports_tools: true.

(b) Add a new tool

Follow experiments/tools/math_tools.py, then register the function and schema in experiments/tools/__init__.py.

def search_tool(query: str, limit: int = 10) -> list[dict]:
    """BM25/vector hybrid search over your knowledge base."""
    ...

TOOL_SCHEMAS.append({
    "type": "function",
    "function": {
        "name": "search_tool",
        "parameters": {...},
    },
})
TOOL_FUNCTIONS["search_tool"] = search_tool

(c) Add a new role

Follow the existing SYSTEM_PROMPT, CRITIC_PROMPT, and JUDGE_PROMPT patterns in experiments/system_prompt.py, then extend the call order using experiments/orchestrator.py:run_abc_chain as the reference.

(d) Add a new taskset item

Append entries to experiments/tasks/taskset.json with id, category, difficulty, prompt, expected_answer, and scoring_keywords. For math tasks, solve the item yourself before trusting expected_answer; Exp08 exposed a broken ground-truth entry in math-04.

Roadmap

Horizon	Item	Status
Phase 1	Exp00~Exp10 — four-axis externalization baseline + cost-aware (Exp10)	✅
Phase 1 follow-up (2026-04-30)	Taskset 3-FAIL fix + Exp09 5-trial drop analysis + Exp10 v3 rescore	✅
Stage 2A/2B (2026-04-30)	Infra stabilization (healthcheck/abort + result JSON meta v1.0) + scorer/failure-label reference	✅
Stage 2C (2026-05-02)	Exp06 H4 recheck on expanded 15-task set — H4 ⚠ Conditionally supported (synthesis +0.140)	✅
Stage 4 (2026-05-03)	Exp11 Mixed Intelligence (Flash Judge) — H10 ⚠ Inconclusive, effectively rejected (inverse mechanism observed)	✅
Stage 5 Exp12 (2026-05-04)	Extractor Role (pre-stage) — H11 ⚠ Conditionally supported (Δ=+0.050, d=+0.323, p=0.198 NS). Directional positive observed	✅
Stage 5 Exp13 (2026-05-05)	Reducer Role (post-stage) — H12 ⚠ Inconclusive, effectively rejected (Δ=−0.053, d=−0.323, p=0.180 NS). Mirrored directional effect to H11	✅
Stage 5 Exp14 (2026-05-05)	Search Tool (agent-active BM25 retrieval) — H13 ⚠ Inconclusive, effectively rejected, statistically significant (Δ=−0.220, d=−1.000, p=0.012 SIG). Mechanism = insufficient retrieval iterations on multi-hop tasks. Tool-axis sub-distinction: deterministic computation (H7/H8 +18~23pp) ≠ agent-iterative retrieval (H13 −22pp)	✅
Stage 6 (planned)	Cross-model replication + LLM-as-judge — H10/H11/H12/H13 generalization on Llama 3.1 8B / Llama 3.3 70B / Qwen 2.5 7B Q4_K_M (Groq free + local). LLM-as-judge to address H12/H13 keyword-scorer caveats. Plan in progress (docs/plans/)	⏳
Mid-term	Remaining unexternalized axes — Search Tool / Graph Tool / Evidence Tool. Stage 5 SQLite ledger
Mid/long-term	Test the four-layer external knowledge environment (vector / graph)
Long-term	Cross-model reproduction on Qwen / Phi / Llama, structured ablations, and a public write-up

How to Contribute

This is currently a 1-person research notebook. Reproduction by anyone running a different model is one of the highest-value contributions.

Difficulty	Example	Contribution form
⭐ 5 min	Typo or documentation fix	PR
⭐⭐ Hours	Reproduce an existing experiment on another model	Issue (Reproduction)
⭐⭐⭐ Days	Add one Tool with tests and an integration run	PR + result report
⭐⭐⭐⭐ Weeks	Design, implement, and evaluate a new Role	PR + research note section
⭐⭐⭐⭐⭐ Months	Long-context stress or cross-model ablation	Acknowledgements top section + possible co-authorship in a future write-up

Contribution flow:

1. Open an issue with one line of intent, for example: I'll try reproducing Exp08 with Qwen 2.5 7B.
2. Fork and run the experiment in your environment.
3. Share code as a PR or results as an issue comment.

Related work

The "externalize the LLM" framing is not unique to this project. Adjacent or overlapping ideas:

• Externalization in LLM Agents (Zhou et al., 2026)¹ — A unified review proposing four externalization axes (memory / skills / protocols / harness engineering). Gemento was developed independently — out of practical context/memory problems hit while building secall and tunaFlow — and only later did the author become aware of this preprint. The four-axis split (Tattoo / Tools / Role / Orchestrator) is best read as independent convergence with that line of work, not as a derivation from it. Axis mapping differs (Gemento separates Role and Orchestrator explicitly; Zhou et al. fold control into harness engineering).
• LightMem (Fang et al., 2026)² — Lightweight memory-augmented generation with three-stage memory (sensory / short-term / long-term, sleep-time consolidation). Focused on long-term retrieval across sessions; Gemento is closer to working state across loops within a single task, not memory across sessions.
• StateFlow (Wu et al., 2024)³ — Conceptualizes complex task-solving as state machines, externalizing control flow from LLM. Conceptually adjacent to Gemento's Orchestrator axis; Gemento adds explicit role separation (A/B/C) and Tattoo schema on top.
• Chain-of-Agents (Zhang et al., 2024)⁴ — Sequential multi-agent reading of long inputs (worker agents segment + manager synthesizes). Gemento's A→B→C pipeline shares this structure but uses the same base model for all roles, separated only by prompt and validation contract.

What is contributed here, and what is not:

• Contributed: a measured workbook of small-LLM (Gemma 4 E4B, effective 4B parameters) behavior across the four axes, with reproducible numbers and sampling parameters.
• Not contributed: a new architecture, a new training method, or a claim that small LLMs replace large ones. Gemento is a structural workflow harness on top of an unmodified open-weight model.

---

¹ Zhou, C., Chai, H., Chen, W., et al. (2026). Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering. arXiv:2604.08224. ² Fang, J., Deng, X., Xu, H., et al. (2026). LightMem: Lightweight and Efficient Memory-Augmented Generation. ICLR 2026. arXiv:2510.18866. ³ Wu, Y., et al. (2024). StateFlow: Enhancing LLM Task-Solving through State-Driven Workflows. arXiv:2403.11322. ⁴ Zhang, Y., Sun, R., Chen, Y., Pfister, T., Zhang, R., Arik, S. Ö. (2024). Chain of Agents: Large Language Models Collaborating on Long-Context Tasks. NeurIPS 2024. arXiv:2406.02818.

Acknowledgements

• Memento (Christopher Nolan, 2000) — original metaphor of external memory aids. Four-axis externalization is the structured-schema cousin of Leonard's tattoos and polaroids.
• secall · tunaflow — the practical origin. The context and memory problems I ran into while building those tools are what made this project necessary.

License

MIT — fork, modify, redistribute, and use commercially. Just keep the copyright notice.