Evolving Memory — AI skill for Claude Code

Evolving Memory

**Cognitive Trajectory Engine (CTE) — Topological memory graphs for LLM agents**

A bio-inspired memory system that captures agent execution traces, consolidates them through dream cycles (SWS/REM/Consolidation), and enables intelligent memory retrieval via topological graph traversal. Built on an **Agentic ISA** (Instruction Set Architecture) where LLMs emit structured opcodes instead of JSON.

[]() []() []() []()

---

The Hypothesis

What is Memory?

LLMs already have memory. They generate content from patterns memorized during training — the weights encode compressed statistical representations of language, reasoning, and world knowledge. In this sense, **memory and learning are inseparable**: what a model "knows" is what it has consolidated from exposure to data.

But this is only one kind of memory. Humans have at least three:

1. Semantic memory — facts and concepts (what LLM weights encode)
2. Episodic memory — specific experiences, ordered in time (what context windows hold temporarily)
3. Procedural memory — skills automated through repetition (what doesn't exist in current LLM systems)

Current LLM agents have strong semantic memory (pre-training) and weak, volatile episodic memory (context window). They have **no procedural memory at all**. Every conversation starts from zero. Every session is independent. There is no mechanism by which an agent improves at a task through repeated practice, learns to avoid past mistakes, or builds upon previous work sessions.

This is the gap Evolving Memory fills.