Sara Brain is a structured knowledge substrate that lets you teach arbitrary facts to an LLM session and measure how faithfully the model retrieves them versus confabulating from training weight. It is also a path-of-thought cognitive architecture: knowledge is stored as directed neuron-segment chains in SQLite, not as model weights.

pip install sara-brain

sara-new --out my_brain.db                                                    # create fresh brain
sara-teach --subject zilkrap --relation is_a --object pruvnet --brain my_brain.db
sara-ask-stateless "what is zilkrap?" --brain my_brain.db                     # fresh session

For training-orthogonal validation substrates (concepts that cannot be in any model's training data):

sara-synth --out synth.db --seed 42    # generates random nonsense-word substrate + manifest

Divergence between B and C is the measurement. Content in B not present in the substrate is confabulation from training weight.

LLM sessions accumulate context. Teaching facts in Session A infects the session context. A truly fresh session (B or C) carries no teaching history — answers can only come from the substrate or model weights. The substrate is the ground truth; the model's answer is the measurement.

DOI: 10.5281/zenodo.19436522

A path-of-thought brain simulation. It learns facts expressed as natural-language statements, stores them as directed neuron-segment chains in SQLite, and recognizes concepts by launching parallel wavefronts from input neurons and finding intersections.

This is not a neural network, LLM, or activation-based system. Recognition follows actual recorded paths through neuron chains — no decay, no forgetting, no pattern matching.

Every essay I publish on Substack (at Path of Thought) or anywhere else is first committed to this repository, in the substack/ directory, before it appears on the public venue. The git history is a timestamped, public record of everything I have written under the Path of Thought name.

If any text appears attributed to me as a Path of Thought essay and is not in this repository, it is not from me. Screenshots can be faked. Substack accounts can be impersonated. AI-generated text can be attributed to anyone. But this repository has been visibly mine for a long time, with consistent work, published research, and real code — it cannot be spoofed by anyone who does not have years of my commit history behind them.

To verify a post: search this repository for a distinctive phrase from the post. If you find it, look at the commit that introduced it — the commit author will be me, and the timestamp will precede the Substack publication date. If you cannot find the text in this repository at all, the text did not come from me.

This authenticity claim applies to written essays under the Path of Thought name. It does not mean every word I have ever written lives in this repo — private conversations, unpublished drafts, and work in other contexts are not required to be here. The claim is narrower and stronger: anything I publish publicly as a Path of Thought essay will be committable-and-committed here first, and anything claimed to be a Path of Thought essay that is not here is not from me.

— Jennifer Pearl

1. Path-of-thought, not activation levels — Recognition traces real recorded paths, not spreading activation
2. Never forgets — Strength only increases. Path similarity replaces forgetting
3. Parallel wavefront propagation — Multiple wavefronts launch simultaneously; commonality across paths determines recognition

python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"

# Run the demo
.venv/bin/python demos/apple_demo.py

# Run tests
.venv/bin/pytest tests/ -v

# Launch the REPL
.venv/bin/sara

Requires Python 3.11+.

• Neurons — four types: concept (apple), property (red), relation (apple_color), association (taste)
• Segments — directed edges between neurons with strength (1 + ln(1 + traversals)) and traversal counts
• Paths — recorded chains of segments representing a learned fact, with source text provenance

Teach it facts in natural language:

sara> teach an apple is red
sara> teach an apple is round
sara> teach a banana is yellow
sara> teach a banana is long

Each statement creates a 3-neuron chain: property → relation → concept.

Give it properties and it finds matching concepts via parallel wavefront intersection:

sara> recognize red, round
Recognized: apple (score: 1.00)

sara> trace apple       # All outgoing paths from a neuron
sara> why apple         # All paths leading to a neuron with provenance
sara> similar apple     # Neurons with shared downstream paths

src/sara_brain/
  models/        — Pure dataclasses (Neuron, Segment, Path, PathTrace, RecognitionResult)
  storage/       — SQLite repos (NeuronRepo, SegmentRepo, PathRepo, AssociationRepo, CategoryRepo, SettingsRepo, Database)
  parsing/       — Statement parser and taxonomy
  core/          — Brain orchestrator, Learner, Recognizer, SimilarityAnalyzer, Perceiver
  nlp/           — VisionObserver (Claude Vision), LLM translator (Claude-only)
  visualization/ — ASCII tree and Graphviz DOT export
  repl/          — Interactive shell with commands and formatters

166 tests covering models, storage, parsing, learning, recognition, similarity, associations, categories, queries, LLM translation, vision, perception, proxy, and end-to-end integration.

.venv/bin/pytest tests/ -v

SQLite with WAL mode and foreign keys. Schema in src/sara_brain/storage/schema.sql. Current plan is SQLite now, with migration to data-nut-squirrel later.

See LICENSE if present.