elfmem

Adaptive memory for LLM agents. Knowledge that evolves through use.

elfmem began as an experiment: could a trading bot develop a concept of self, learning which strategies succeeded, which failed, and evolving its approach through experimentation? That question led to a fundamental insight: agents don't need a database of facts. They need memory that adapts. Knowledge that gets reinforced through successful use should grow stronger. Knowledge that misleads should fade. And the agent's identity, its values, style, and hard-won lessons, should persist across every session.

We built elfmem from the ground up to make this real. It's a memory system modelled on how biological memory works: fast ingestion, deep consolidation at pauses, adaptive decay at rest, and a knowledge graph where related ideas strengthen each other over time. Every design decision was derived from first principles across 26 structured explorations, not borrowed from existing patterns, but built from axioms about how agent memory should work.

One SQLite file. Zero infrastructure. Any LLM provider.

An agent's knowledge after several sessions. Nodes are memory blocks, sized by confidence and coloured by decay tier. Edges are semantic relationships discovered during consolidation. Identity blocks (permanent tier) anchor the centre. Knowledge that gets used grows; knowledge that doesn't fades toward the periphery.