primordial-soup-

中文 | English

The first collision-based memory system: it doesn't retrieve old ideas — it smashes old cards into new ones.

A cross-conversation association mechanism: conversations crystallize "amino-acid cards"; each new session, a "lightning strike" — a Miller-Urey-style random collision — smashes randomly drawn old cards into today's topic.

Positioning: as of Aug 2026 research, no public project combines these four layers. The mainstream memory routes are retrieval (semantic search) and compression (summarization). This mechanism's core action is collision: randomly flip up old cards and crash them into the current topic, letting unexpected connections produce new ideas. Public neighbors — the A-MEM line (graph linking + semantic retrieval + LLM-driven evolution) and the memory-compiler line (summarization + embedding retrieval) — cover none of the four-layer stack: true-random sampling (with a wildcard slot) × weight evolution (Hebbian / Zeigarnik / decay) × protein synthesis × full ledger audit. Plenty of projects retrieve memories. Almost none collide them.

原始汤 (primordial soup, from the origin-of-life metaphor) is an association system that keeps memory in the filesystem: each conversation's conclusions crystallize into an "amino acid" card; at the start of a new session, a "lightning strike" randomly draws a few old cards and smashes them into today's topic — the sparks may correct old conclusions, answer old open questions, or synthesize new "protein" review cards. All memory lives in files, with no dependency on any global memory.

• Amino acid ＝ one conversation's conclusion card (one strong claim per line, plus open questions and connection points)
• Protein ＝ a synthesis of several cards (joins the pool, but never co-drawn with its members)
• Lightning ＝ weighted random draw (2 weighted + 1 wildcard), smashing old cards into today's topic
• Tide ＝ a weekly reconciliation task: collect pending verdicts, detect clusters, propose closures (proposes only, never implements)

Features

• True-random collision: card selection is computed deterministically by the sampler (system entropy, or --seed for reproducibility). The model is forbidden to hand-pick cards — deliberate picking kills the very serendipity lightning exists for.
• Incremental presentation gate: a collision is presented only if it passes at least one of four incremental tests (changes a conclusion / provides a counterexample or tension / answers an old open question / opens a genuinely new direction). Otherwise total silence — no mention, no "no collision this time" ceremony.
• Silence is auditable: even silent draws are logged. The silence rate feeds the M5 effectiveness settlement, so gate tightness is measurable and tunable.
• Verdict power belongs to the user: the AI only proposes; the user clicks the verdict (strong-accept / accept / indifferent / reject, one card per question). Strong-accept triggers the Hebbian ×2 weight, and must point at a concrete location (structural criterion — intensity adjectives are banned).
• Weights are script-computed, never model-computed: Hebbian learning, Zeigarnik keeping-warm, and forgetting decay are all deterministic script math.
• Append-only: card numbers never change once issued; bidirectional write-backs only append annotations — old text that cited a number can never point at the wrong card.
• Host-agnostic: the mechanism spec and scripts are plain files + command line; host-specific adaptations (e.g. for DSH) live in their own layer under adapters/.

Five-minute quickstart

See examples/GETTING-STARTED.md (Chinese): copy the synthetic mini-soup → lightning → settle a card → status page. Fully synthetic data, whole loop runs in minutes.

Layout

primordial-soup\
  README.md / README.en.md   this file, bilingual
  LICENSE                    Apache-2.0
  spec\
    SOUP-MECHANISM.md        the mechanism spec (host-agnostic: range guard / folder contract / numbering / tag dimensions / lightning / settlement / sentinels / fast-track closure / maintenance)
    lightning-trigger.md     trigger design T1-T6 (when lightning fires, dedup criteria, settlement hooks, boundary table)
  scripts\
    闪电抽样器.py            M3 weighted sampling + wildcard slot (read-only, zero side effects)
    状态页生成器.py          generates _氨基酸库\状态页.html (pre-generated static page)
    M5结算脚本.py            effectiveness settlement (closed questions / strong-accepts / silence rate)
    潮汐周报生成器.py        weekly report md → HTML (with content reconciliation; exit 3 = missing sections)
  examples\
    mini-soup\               synthetic sample soup (all cards are fictional test data — copy and remix freely)
    GETTING-STARTED.md       quickstart walkthrough
  adapters\
    ds-harness\              adapter for a specific host (SKILL.md, INSTALL.md, porting notes)

Script parameters (quick reference)

The weight formula structure, the four incremental gate criteria, silence-auditability, and append-only write-backs are mechanism invariants — deliberately not parameterized.

Design principles

1. True randomness: only the sampler may be random (entropy or --seed); the model must not pick cards.
2. Presentation gate: fail the incremental test and stay completely silent — no mention, no explanation, no ritual trace.
3. Auditable silence: silent draws are logged anyway; the silence rate feeds M5 settlement, so gate tightness can be calibrated with data.
4. User holds the verdict: AI recommendations are proposals; the click is the verdict. Strong-accept must cite a concrete location (structural criterion, no intensity adjectives).
5. Script math, model no-discretion: Hebbian ×2 / Zeigarnik keeping-warm / forgetting decay are all computed deterministically by scripts.

Adapters

The spec and scripts are host-agnostic; adaptations to a specific host (agent runtime / workspace conventions / interaction tools / scheduling) live under adapters/:

• adapters/ds-harness/ — for the DSH host: SKILL.md (full skill), INSTALL.md (install locations / hot reload / two script-distribution modes), PORTING-NOTES.md (porting checklist + design notes: dependency mapping, script-level findings, contract baselines, R1-R7 work orders).

Provenance

The mechanism and scripts were battle-tested in a private deployment before being open-sourced. This public repository contains no private data — the sample soup (examples/mini-soup/) is entirely synthetic and fictional.

License

Apache-2.0, Copyright 2026 primordial-soup contributors.