Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.4stax.com/llms.txt

Use this file to discover all available pages before exploring further.

kontxt ranks memory candidates using four signals: semantic similarity, recency, access frequency, and explicit importance.

​
Scoring model

At a high level: 
score = (semantic_similarity × 0.50)
      + (recency_decay       × 0.20)
      + (access_frequency    × 0.15)
      + (importance_score    × 0.15)
These weights are meant to reflect a simple principle. Semantic similarity matters most. The other signals help keep results practical. Signal details:
  1. Semantic similarity: embedding similarity between the current query and candidate memories.
  2. Recency decay: exponential decay with roughly a 30 day horizon.
  3. Access frequency: log scaled. Memories that are repeatedly useful get a boost.
  4. Importance score: explicit weight set at write time.

​
Why recency and frequency exist

Semantic similarity alone is not enough. Two memories can be semantically similar but one may be outdated. Recency and frequency help kontxt prefer memories that have proven useful recently.

​
Importance

Importance is a manual override. Use it sparingly and only for items you always want to retrieve, such as core preferences and irreversible architectural decisions.

​
Embedding tiers (no mixing)

Memories are tagged with an embedding_tier. Similarity is only computed among memories from the same tier so switching providers doesn’t produce meaningless scores. Supported tiers (in priority order):
  1. OpenAI (text-embedding-3-small). Optional key.
  2. Transformers.js (all-MiniLM-L6-v2). Offline after first model download to ~/.kontxt/models/.
  3. Ollama. Used when ollama serve is running with an embedding model pulled.
  4. Pseudo. Keyword level fallback. Always works.
See Core concepts for the mental model, and the kontxt pages for integration surfaces.