Extract the text between *marker*: and the next known marker (or end). Args: text: Full LLM response marker: Section header to fi

Parse a section of text into a list of items. Handles: - Bullet points (-, *, numbered) - Comma-separated values - One item per

Search for similar documents using cosine similarity. Args: query: Query text k: Number of results to return

Load retriever state from disk

One-time Add documents to both BM25 and semantic index

Remove ```json ... ``` or ``` ... ``` fences from LLM output.

Add a new memory note

Extract heuristic keywords from content text.

Load the LoComo dataset from a JSON file, including image-based content by using captions. Args: file_path: Path to the JSON fil

Try JSON parsing first; fall back to section-marker parsing. Many models emit valid JSON even without strict mode, so we try that first for b

Save retriever state to disk

Extract a heuristic context sentence from content.

Calculate aggregate statistics for all metrics, split by category.

Calculate comprehensive evaluation metrics for a prediction.

Find related memories using hybrid retrieval

Find related memories using hybrid retrieval

Load retriever state from memory

Simple tokenization function.

Validate and repair the analysis result. - If keywords is empty, extract capitalized words / nouns heuristically. - If context is empty, use

Analyze content using plain-text prompt + section-marker parsing.

Analyze content to extract keywords, context, and other metadata

Generate answer for a question given the conversation context.

Generate answer for a question — plain text, no JSON schema.

Calculate BERTScore for semantic similarity.

Calculate BLEU scores with different n-gram settings.

Calculate METEOR score for the prediction.

Calculate ROUGE scores for prediction against reference.

Calculate sentence embedding similarity using SentenceBERT.

Send a test call to verify the backend is reachable.

Re-initialize the retriever with current memory state.

Consolidate memories: update retriever with new documents This function re-initializes the retriever and updates it with all memory d

Evaluate the agent on the LoComo dataset. Args: dataset_path: Path to the dataset file model: Name of the model to use

Evaluate the robust agent on the LoComo dataset.

Find related memories using embedding retrieval.

Generate query keywords — plain text, no JSON schema.

Get a plain-text completion from the LLM.

Parse the analyze_content LLM response. Returns: {"keywords": [...], "context": "...", "tags": [...]}

Parse conversation data.

Parse the evolution decision response. Returns: {"decision": "NO_EVOLUTION|STRENGTHEN|UPDATE_NEIGHBOR|STRENGTHEN_AND_UPDATE", "r

Parse generate_query_llm response. If JSON with "keywords", extract it. Otherwise return raw text.

Parse a plain-text answer response (for QA evaluation). If the model returned JSON with an "answer" field, extract it. Otherwise return the r

Parse retrieve_memory_llm response. If JSON with "relevant_parts", extract it. Otherwise return raw text.

Parse a single session's data, including turns with images by using their captions.

Parse the strengthen details response. Returns: {"connections": [int, ...], "tags": [str, ...]}

Parse the update neighbors response. Returns: [{"context": "...", "tags": [...]}, ...] — one per neighbor

Process a memory note for evolution using plain-text LLM calls. Uses up to 3 sequential calls (conditional): 1. Evolution decision

Process a memory note and return an evolution label

Run system tests

Set up logging configuration.

Set up logging configuration.

Initialize the hybrid retriever. Args: model_name: Name of the SentenceTransformer model to use alpha: Weight

Initialize the simple embedding retriever. Args: model_name: Name of the SentenceTransformer model to use

Add a single document to the retriever. Args: document: Text content to add Returns: boo

Add documents to the retriever.

Add a new memory note.

Get the appropriate answer based on category.

Find related memories with neighborhood expansion.

Get completion from LLM