wasnaga/atom-token-llm-v1

Punk Records Atom Token LLM (v3.0-alpha)

Separable memory for language agents — session state in a neural ψ field, not transcript token replay.

Paper	Full paper + figures (PDF) · Research overview · Evidence pack
Live demo	wasnaga/atom-token-demo
Author	Umberto Canessa Cerchi · umberto.canessa@gmail.com
Stack	~3.5M params · v3.0.0-alpha · release + v3 closure gates PASS (100%) · 2026-07-07

---

Abstract

Chat agents remember by re-tokenizing prior turns. Cost grows with session length; latent multi-agent systems share opaque KV caches that resist audit.

Atom Token LLM uses a native unit — the AtomToken (σ, ψ, π, φ) — and a ψ field of content-addressable cells (κ, ν). Surface text appears only at RENDER egress. A learned tension field routes writes and reads; a ψ-predictor enables imagination-based plan_act without executing destructive actions.

On deterministic multi-agent benchmarks (500 TEST societies, held-out seeds):

• Portable mentorship: ψ-transfer 500/500 vs control 0/500
• Latent guardrail: post-lock deploy refused 500/500
• Stale-belief rent: planner 0% stale redeploy vs reactive 100%
• Dream quarantine: 0 violations
• Routing @ banks 8–64: 100% routing recall and read-given-routed
• Horizon recall: 100% @ 8/16/32 facts; 97.5% token savings vs transcript @ 32 facts
• Society input savings: 31%
• v3 STUDY / NL induction / conduct gates: PASS on TEST seeds (Act 5)

Full write-up: docs/RESEARCH.md.

---

Why this exists

Mainstream chat LLMs remember by replaying the conversation as tokens every turn. Atom Token LLM inverts the stack: each inference step takes only the latest user text + ψ state.

Layer	Symbol	Role
Index	κ = (κ_coarse, κ_fine)	Learned hierarchical address
Substance	ν	Neural value — the fact as stored
Surface	text	RENDER egress only

Product invariant: no forward pass over prior chat as BPE token IDs. Session evolution is lawful ψ transition; weights provide ability, ψ accumulates experience.

---

v3.0-alpha — emergent procedures and conduct

v2 proved emergent realms (facts the model was never told exist). v3 proves emergent procedures and conduct — skills and behavioral rules taught through natural language, executed from ψ at inference, with no post-ship training.

Layer	What ships	Proof
STUDY (C′)	Curriculum via send() → proc_* cells in episodic ψ	G-skill-study-*
CONDUCT (D′)	Policy constraints via send() → policy_* cells	G-behavior-study-*
NL induction (F)	Atomizer v7 decomposes novel teaching paraphrases	G-nl-induction-*
Skill field	Verified traces → composites; fluency, rust, .atom transfer	G-skill-fluency/rust/transfer
Unified codec	One value_codec.pt for facts + proc/policy ν (v35c)	3942/3942 + 12/12 induction

Constitutional law: weights trained once pre-ship; teaching, practice, and fluency are writes only.

---

Architecture (v2.3 substrate + v3 heads)

Component	Role
Atomizer v4	(π, spans) — release routing (society gates)
Atomizer v7 NL	NL induction decomposition (proc_*, policy_*)
κ encoder + κ skill	Hierarchical routing; proc/policy contrastive rows
Tension field v4	WriteMode on PLANT; pull-first QUERY
ValueCodec	ν encode/decode + re-encode verified search (unified facts + induction)
APPLY interpreter	Step loop over planted proc_* cells; skill-field probe first
ψ-predictor	Latent transition for plan_act + composition proposals
Unified ψ-catalog + CRS	Chitchat without transcript replay

Hybrid memory retrieval (Findings #9–#11):

1. κ_coarse cascade — bounded window k=8 over unbounded bank
2. Structural write/read laws — ν-matched UPDATE, cross-facet veto
3. Lexical anchor — exact token match on OOD facets when learned read disagrees
4. Value-shape law — IPv4 as ip:<addr>; deploy grounding veto in plan_act

Persistence: portable .atom file (cells + policy + atom log).

---

Results at a glance (TEST seeds)

Society (Act 4 · seeds 4500–4999)

Claim	Metric
Completion (reactive / planner / dreamer)	500/500 each
ψ-transfer mentorship	500/500 · control 0/500
Fact diffusion	500/500 · mean latency 1.0 rounds
Post-lock deploy refused	500/500
Stale redeploy after unobserved lock	planner 0% · reactive 100%
Dream quarantine	0 violations
Transcript arm (visibility ablation)	0/500
Token savings	31.0% (36,500 / 117,687)

Routing cascade (G-routing)

Bank	Window	Routing recall	Read | routed
8	8	40/40 (100%)	40/40 (100%)
16	8	80/80 (100%)	80/80 (100%)
32	8	160/160 (100%)	160/160 (100%)
64	8	320/320 (100%)	320/320 (100%)

Horizon (Act 1 · post v33 codec)

N facts	Recall	Baseline tok	Kernel tok	Savings
8	8/8 (100%)	7,140	694	90.3%
16	16/16 (100%)	26,969	1,376	94.9%
32	32/32 (100%)	119,542	3,041	97.5%

Codec & plan_act

Gate	Result
G-codec-fact-exhaustive-pool	3942/3942 (100%)
G-codec-horizon-roundtrip	18733/18733 (100%)
G-codec-induction-substance	12/12 (100%)
G-codec-society-deploy-roundtrip	1067/1067 (100%)
G-plan-act-pick	500/500
G-plan-act-refuse	500/500
G-wm-* (outcome, manifold, rollout, counterfactual)	100%

v3 closure (Act 5 · C′ + D′ + F + F′)

Gate	Result
G-skill-study-mul10	50/50
G-skill-study-chain	50/50
G-skill-study-negative	0/20 (control)
G-skill-fluency / rust / transfer / imagine	PASS
G-behavior-study-plant	97/97
G-behavior-transfer	97/97
G-nl-induction-skill-mul10	50/50
G-nl-induction-behavior	97/97
G-nl-induction-transfer	97/97

Template parsers off on default path. Full Act 5 JSON: evidence/act5_v3_closure.json.

---

Download weights

pip install -U "huggingface_hub[cli]" torch pyyaml numpy
huggingface-cli download wasnaga/atom-token-llm-v1 --local-dir ./checkpoints/lane_d

Set LANE_D_CHECKPOINT_DIR to the download directory before running gates locally.

---

Checkpoints in this repo

File	Role
atomizer_v4.pt	Required — release π + multi-slot (society gates)
atomizer_v7_nl.pt	NL induction atomizer (skill + conduct gates)
kappa_encoder.pt	Required — hierarchical κ (release)
kappa_encoder_skill.pt	Proc/policy κ rows
tension_field_v4.pt	Required — memory read/write routing
value_codec.pt	Unified ν codec v35c (facts + induction)
sigma_encoder.pt	Query↔plant σ binding
psi_predictor.pt	World model / plan_act
unified_chitchat_v16.pt	Chitchat primary — ψ-catalog + CRS
stack.json	Manifest (v3.0.0-alpha)

See stack.json for full manifest.

---

Reproduce eval

From punk-records-atom-token with pinned checkpoints:

export LANE_D_CHECKPOINT_DIR=/path/to/checkpoints/lane_d
export LANE_D_ALLOW_HF_DOWNLOAD=0
export PYTHONPATH=.

python -u -m lane_d.product.eval.run_release_gates      # ~22 min — codec, routing, wm, society
python -u -m lane_d.product.eval.run_fp_closure_gates   # ~5 min — v3 C′/D′/F + skill-field
python -u -m lane_d.society.eval.export_evidence          # ~25 min — Acts 1–5 paper tables

# Individual v3 gates
LANE_D_NL_INDUCTION_GATES=1 python -u -m lane_d.product.eval.g_nl_induction
LANE_D_SKILL_GATES=1 python -u -m lane_d.product.eval.g_skill
LANE_D_BEHAVIOR_GATES=1 python -u -m lane_d.product.eval.g_behavior

---

Citation

If you use this work, please cite:

Canessa Cerchi, U. (2026). Separable Memory for Language Agents: ψ Fields, Latent World Models, and Multi-Agent Diffusion (Draft v1.2).
Paper: https://huggingface.co/wasnaga/atom-token-llm-v1/blob/main/paper/atom-token-llm-full-paper.pdf
Model: https://huggingface.co/wasnaga/atom-token-llm-v1 · Demo: https://huggingface.co/spaces/wasnaga/atom-token-demo

BibTeX:

@misc{canessa2026atomtoken,
  author       = {Canessa Cerchi, Umberto},
  title        = {Separable Memory for Language Agents},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/wasnaga/atom-token-llm-v1},
  note         = {Draft v1.2; v3.0.0-alpha weights; PDF in repo /paper/}
}

---

Limitations

Deterministic ops deploy simulator · primarily 2-agent societies · ~3.5M parameter stack. Not a frontier general LLM. v3 alpha proves study/induction at template-scale curricula; Phase E generalizes to free text and larger encoders. See paper §9–§10.