Apppublic

barissozudogru/cortex-ai-primitives

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes

App README

Cortex AI - cognitive primitives for LLM agents

Three live demos of the cognitive primitives that move an agent's context layer beyond a filing-cabinet store. Each tab is also exposed as an MCP tool so an LLM agent can call the same logic directly.

The algorithms are ported from cortex-ai, a proactive MCP-server context engine for macOS. The live system reads from a local SQLite database populated by connectors that watch calendar, screen, git, shell, and browser activity. This Space substitutes a curated synthetic dataset so the primitives can be exercised in a public sandbox.

What's in here

Tab	Question it answers
Forgetting curves	What does the user still remember, and when does each fact fall below useful recall?
Negative-space retrieval	What entities historically co-occur with the current work frame but are absent right now?
Hippocampal consolidation	Compress this stream of episodes into a smaller set of semantic capsules.

Why this matters for agents

A long-context window is not memory. Real cognitive memory has structure: it decays at predictable rates, it surfaces the most salient gap when reasoning about a familiar pattern, and it consolidates fine-grained moments into reusable abstractions during idle periods. An agent that treats its context as a flat lookup table inherits none of those properties.

Each primitive here is a deliberate slice of human cognition retargeted at machine context:

—Forgetting curves prevent re-surfacing facts the user just saw, and identify staleness before retrieval quality drops. Anderson (1991) ACT-R activation, log-exposure stability per Wozniak/Anderson.
—Negative-space retrieval turns absences into evidence. If JWT and refresh_token are active but CSRF is missing despite a 0.55 historical co-occurrence, the agent should ask why.
—Consolidation mirrors hippocampal replay: episodic events get distilled into reusable semantic capsules so downstream retrieval reads a few capsules instead of thousands of raw events.

MCP usage

Add this Space to your MCP client config:

json

{
  "mcpServers": {
    "cortex-primitives": {
      "url": "https://barissozudogru-cortex-ai-primitives.hf.space/gradio_api/mcp/sse"
    }
  }
}

Three tools become available:

—compute_forgetting(items, horizon_days) - project retention probability over a horizon for a list of items.
—find_negative_space(active_entities, min_probability, min_significance, limit) - surface ranked ghost entities for a given work frame.
—consolidate_episodes(episodes, distance_threshold) - cluster episodes into semantic capsules.

Cold start

This Space sleeps when idle. First request after a quiet period takes 30-60 seconds for the container to wake up. Subsequent requests are fast.

Citations

—Anderson, J. R. (1991). The Adaptive Nature of Human Categorization. Psychological Review 98(3).
—Luhn, H. P. (1958). The automatic creation of literature abstracts. IBM Journal of Research and Development 2(2).
—Wozniak, P. A. (1990). Optimization of learning. Repetition spacing PhD foundation.
—Ebbinghaus, H. (1885). Über das Gedächtnis. (English translation 1913, Teachers College Press.)
—Agrawal, R. & Srikant, R. (1994). Fast algorithms for mining association rules. VLDB.