A biologically-inspired, neuro-symbolic cognitive architecture for lifelong learning. Digital neuromodulators regulate learning, hypothesis tracking, and memory consolidation over an SQLite backend. An autonomous, biologically inspired AI system designed to learn to model the underlying mechanics, structures, and rules of language and text—rather than just memorizing facts.
Code export on NotebookLM, for Code review
Traditional AI models and semantic databases often focus on memorizing and reproducing content (the "what"). Brainstem takes a fundamentally different approach (the "how"). The system treats incoming data (such as massive text corpora or Wikipedia dumps) not as a static knowledge base to be memorized, but as a training substrate. Its goal is to analyze, comprehend, and model the structural and semantic grammar of language.
- Structural Learning: Brainstem learns how subjects, relations, and objects interact, mapping how meaning is built dynamically.
- Noise Filtering via Sleep Cycles: By mimicking biological sleep phases (such as slow-wave sleep and consolidation), the system constantly prunes weak or chaotic connections to crystalize the core rules of language.
brainstem (Test Version 8) does not store data statically. Instead it simulates the dynamics of a biological brainstem: a homeostatic network of 12 digital neurotransmitters regulates data-processing rates, evaluates uncertainty, forms context hypotheses, and consolidates knowledge during simulated sleep cycles.
The system extracts logical relations from unstructured text and incrementally builds a probabilistic world model in a relational SQLite database. Learning rate, exploration, error weighting, consolidation, and sleep–wake regulation are controlled autonomously — the system is designed to learn how to learn, not to be told the rules.
| Stage | Name | Description |
|---|---|---|
| 1 | Inference-Free Pre-Parsing | The raw corpus (e.g. Wikipedia .zim) is fully read, structured, and partitioned into a high-performance chunk store before learning begins (GUI: "ZIM Einlesen"). |
| 2 | Autonomous Learning | The AutonomousLoop processes prepared chunks in bite-sized portions, so neuromodulation reacts purely to data semantics — undisturbed by disk I/O latency. |
Phases are layered sequentially (4x → 5x → 6a–6d → 7a–7g → 7cort). Loading is handled by a single declarative registry (ki_system/phase_registry.py):
- Defines the exact load order in one place
- Isolates each entry with its own error handling + machine-readable load report
- Runs a post-load self-check (verifies the cycle chain terminates at the correct top phase and that all safety flags are set)
Processing chain (top → down):
Cortisol (7cort) → 7g BDNF → 7f Orexin → 7e Histamine → 7d Slow-Wave → 7c → 7b1 → 7b → 7a → 6d → 6c → 6b → 6a
Each cycle computes 12 self-regulating chemical messengers. Every level is derived from the system's own internal state (progress, uncertainty, sleep pressure, consolidation consistency, reading coverage) via data-dependent dynamics on top of fixed biophysical constants (hard bounds [0.0, 1.0], decay rates, homeostatic set-points) — analogous to how biology combines fixed physiology with adaptive regulation. All values are normalized to [0.0, 1.0].
| # | Neurotransmitter | System Role |
|---|---|---|
| 1 | Dopamine (DA) | Success & reward metric — signals gap closure, stabilizes validated pathways. |
| 2 | Serotonin (5-HT) | Consolidation & stability — protects established knowledge from being overwritten (consolidation bias). |
| 3 | Glutamate (Glu) | Excitatory exploration — drives new-hypothesis generation, raises plasticity (exploration bias). |
| 4 | GABA | Inhibitory noise filter — dampens redundant pathways, prevents saturation; reciprocal Glu–GABA E/I balance. |
| 5 | Acetylcholine (ACh) | Novelty detector — controls attention & structural revision when data deviates from the model (revision bias). |
| 6 | Noradrenaline (NA) | Stress & error signal — reacts to chronic blockage (persistent_pressure), forces strategy switches. |
| # | Neurotransmitter | System Role |
|---|---|---|
| 7 | Adenosine phase7a |
Sleep-pressure homeostat — accumulates while awake, forces sleep-replay at max, depletes after replay. |
| 8 | Endocannabinoids phase7b |
Retrograde gain control — 2-AG / Anandamide dampening & LTD of over-excited pathways. |
| # | Neurotransmitter | System Role |
|---|---|---|
| 9 | Cortisol / HPA phase7cort |
Global stability watcher & regulator. Measures allostatic load (threshold drift, survivor collapse, effectiveness depression, oscillation, E/I saturation). Runs as a Stage-2 active regulator at the top of the cycle chain: applies small, capped neuromodulator nudges (±0.05 per cycle, with cooldown, self-limiting) when allostatic load is high, and stays silent when the system is calm. |
| 10 | Histamine phase7e |
Wake / arousal — reciprocal antagonist of adenosine; together they form the emergent sleep–wake switch. |
| 11 | Orexin phase7f |
Reading-endurance & curiosity drive — motivates continued reading while unread corpus and progress remain; self-attenuates when satiated (curious_drive / balanced / satiated). |
| 12 | BDNF phase7g |
Activity-dependent growth & consolidation substrate — rises on consistent, progressing consolidation (growth), else maintenance / low_plasticity. |
In artificial sleep, excitation (glutamate) is throttled while inhibition (GABA) and consolidation (serotonin) dominate. Before hypotheses are strengthened, they pass the _critic_gate:
- Computes
anchor_consistencyof new relations against stable anchor data. - On contradiction/instability →
critic_rejected+ mathematical penalties (hallucination protection). - After a successful replay cycle,
adenosine_levelis fully depleted.
Adds a true <1 Hz up/down-state substructure with self-regulating down-selection:
- Stochastic reactivation (weighted sampling without replacement) — participation varies across oscillations, so consistency can be measured; only repeatedly, robustly active hypotheses survive.
- Adaptive selection threshold derived from the system's own activity distribution and its GABA/Glutamate-derived selection pressure — no hard-coded cut-off.
- Anchor interleaving as an anti-hallucination reality check.
This replaces the earlier "reinforce-everything" no-op with genuine, biologically-plausible consolidation (a small survivor fraction is reinforced, the majority weakened).
Because the knowledge base grows continuously, a marginal progress metric assesses learning honestly (the raw global average is diluted by accumulated mass). Comparing newest-vs-oldest cohorts reveals the true learning curve:
- Uncertainty falls from ~0.99 (earliest hypotheses) toward ~0.68 (most recent)
- The share of undifferentiated
uncertain_hypothesisroles decreases over time
Old, unresolved hypotheses are never discarded — they form a growing review pool, re-examined as more context is read.
Core principle: nothing is thrown away — everything stays open for re-evaluation as knowledge grows.
- Own Test corpus: Wikipedia category "Computer" (German) (~102k chunks, fully pre-parsed as a ZIM import). You must bring your own ZIM file.
- Reading coverage: Learning is corpus-bounded; observed plateaus correspond to not-yet-read material rather than true stagnation. Coverage is tracked explicitly and rises as cycles run.
- Safety mode (Stage A lock): During calibration of phases 6 & 7, permanent writes (
fact_promotion,direct_fact_writes,direct_relation_writes) are explicitly disabled, and no word blacklists are used in the active learning path (legacy phase3d extraction filters are inactive / being removed). Learning happens purely in transient metaplasticity/hypothesis states to verify stability and convergence over hundreds of cycles. - Roadmap: Stage A (stabilize the learning core) is complete and its stability has been empirically verified via the sensory-deprivation drift test (no harmful drift under input removal). Stage B — controlled, consolidation-gated opening of the write locks (with a warm-up dampening of the first promotions to avoid overfitting on the earliest references) — is the next major milestone.
- Future: switch to a vector database.
The system uses a single SQLite database (ki_memory.sqlite3) in the project root.
No manual setup is required. If the database file does not exist on startup, it is created automatically:
- The core schema (documents, chunks, hypotheses, settings, etc.) is initialized on first launch.
- Every phase module carries its own schema and applies it idempotently via
ensure_schema()at the start of each cycle — missing tables are created and missing columns are added automatically (self-healing schema). - Performance indexes are ensured on every start (idempotent), and a built-in
_self_check_schema()verifies all required columns exist before any writes occur.
As a result, you can simply run the GUI on a clean checkout; the database and all required tables/columns are provisioned on demand.
For calibration and repeatable experiments, the learning state can be reset without re-importing the corpus. Available as a GUI button in Export/Konfig ("Lernsystem zuruecksetzen (Korpus bleibt)") and as a standalone script (reset_learning.py).
- KEEP (preserved): chunks, chunks_fts*, documents, settings/config, import_state, sqlite internals — i.e. the full imported ZIM corpus stays intact.
- WIPE (cleared): all learning-generated tables (context_hypotheses, chunk_attention_scores, all phase* state/event/cycle tables, cortisol/neuromodulator states, reading queue, gaps, etc.).
- Safety: dry-run by default (
python reset_learning.pyshows the plan, changes nothing); the real reset (--applyor the GUI button) makes an automatic timestamped DB backup first, then DELETEs rows (tables/schema are never dropped) and VACUUMs. - GUI guard: the reset button is only active while autonomous learning is stopped, and asks for confirmation.
This lets the system be returned to a clean "just-imported" state in seconds, after which it re-seeds all neurotransmitters to their baselines on the first cycle and starts learning from zero.
A GUI-controlled sensory-deprivation mode (tab "Drift-Report") for homeostatic calibration: the wake/read step is skipped (no new chunks, no new hypotheses) while the inner dynamics — sleep replay, consolidation and all 12 neuromodulators — keep running. This isolates the system's self-regulation from input noise and answers a single question: does anything drift when no input arrives?
- Controls: Start/Stop, plus an optional cycle-limit field (up to 9999). With the limit checkbox off, the run continues until Stop; with it on, it stops after N cycles — Stop always remains effective.
- Per-cycle CSV log: every deprivation cycle is written to
drift_log_<timestamp>.csvin the project root (all 12 neurotransmitters + exploration_bias, plasticity, adaptive_threshold, survivors/participated/weakened, effectiveness, reciprocal_gate, allostatic_load). - Live graphs: key signals (exploration_bias, adenosine, plasticity, effectiveness, histamine) are plotted live on a canvas.
- Drift report: for each signal — first/last/delta/span and a verdict (stabil / konvergiert / DIVERGIERT); count-signals (survivors etc.) are judged relative to their mean, normalized signals by absolute thresholds. Overall verdict: kein_drift / konvergenz / DIVERGENZ-WARNUNG. Appended to the CSV as a comment block.
- Fail-safe: the deprivation flag is always cleared on stop/finish (finally), so the system can never remain stuck in deprivation.
Result so far: over 1300 input-free cycles the regulated signals (neuromodulators, exploration_bias, plasticity) stay within a very tight band — the system is homeostatically stable and does not drift under input removal.
The GUI is launched from the project root:
python main.py --gui| Step | GUI Action | Purpose |
|---|---|---|
| 1 | Export/Konfig | Set the maximum number of articles for the ZIM import (must be configured before importing). |
| 2 | Import & Jobs → ZIM Einlesen | Stage 1 — pre-parse & import the corpus into the chunk store (run once). |
| 3 | Import & Jobs → Autonom dauerhaft starten | Stage 2 — start the autonomous learning loop; each GUI cycle runs several internal chain passes. |
| 4 | Import & Jobs → Autonom stoppen | Stop the loop cleanly. |
| 5 | Possible database damage. |
Note on the GUI: This is an experimental testing interface. Currently the Import & Jobs, Export/Konfig and Drift-Report tabs are functional — the remaining tabs are placeholders. Import & Jobs shows a live 12-neurotransmitter dashboard (with tooltips and legend), two progress bars (total corpus coverage + current GUI-cycle step), and a small floating "mood head" window that mirrors the system state (curious / growing / sleeping / stressed) and auto-closes with the GUI.
The maximum article count for the ZIM import is configured in Export/Konfig and must be set before running ZIM Einlesen.
Required: zimdump for Windows (the .exe together with its required .dll files) must be placed in the project root directory. Without it, the ZIM Einlesen function will not work.
The ZIM import relies on zimdump to extract articles from the .zim archive. Make sure the executable and all accompanying DLLs sit next to main.py in the root folder before running Stage 1.
- Codebase language: Documentation is English; the source code (comments & internal naming) is written in German.
- AI-assisted engineering: Developed with the collaborative assistance of advanced language models (Claude 3.5/4 Opus, ChatGPT, Gemini, NoteBookLM).
- Delivery & verification workflow: Every module ships with idempotent schema management, self-checks, and a smoke test before integration; patches are validated with compile checks, backups, and automatic rollback.
- Project scale & status: A massive, highly experimental testing system. Due to its scale and ongoing calibration of complex homeostatic loops, the architecture has not yet been consolidated — it remains a playground for heavy benchmarking and mathematical verification of neuromorphic concepts.