Ferricula

Thermodynamic memory engine for AI agents.

Overview

Ferricula is a persistent memory engine designed for AI agent fleets. It stores memories as entries with fidelity scores that decay over time, consolidates related memories through dream cycles, and exposes a fast query API via HTTP on :8765.

The source is closed. The binary is public. Pull it and run it.

docker run -p 8765:8765 -v ferricula-data:/data kord/ferricula

Memory Model

Fidelity Scores

Every memory carries a fidelity score (f32, 0.0–1.0). The score decays over time at a rate determined by access patterns — frequently recalled memories decay slowly, neglected ones decay faster. When fidelity drops below the survival gate, the memory transitions state.

Cognitive Heat

Each agent identity accumulates cognitive_heat during rapid recall. When heat exceeds the Fortune gate threshold, the engine temporarily stops surfacing memories — protecting against recall feedback loops. Heat dissipates passively or through a dream cycle.

Dream Cycles

An optional consolidation process that clusters related memories, discovers semantic structure, and promotes high-value memories to keystone status. Keystoned memories stop decaying entirely. Dream cycles also actively cool cognitive_heat.

Channel Isolation

Each agent writes into a named channel. Memory queries are scoped to a channel — agents in a fleet don't see each other's context unless explicitly cross-queried.

Query Interface

SQL-style API

Ferricula exposes a SQL-style query language. SELECT memories with WHERE clauses, tag filtering, and tag_jaccard() similarity scoring. Agents and tools can query memory with the same syntax as any structured store.

Bitmap Index

The query engine uses bit-sliced RoaringBitmaps. Tag intersections, channel scans, and fidelity range filters resolve in microseconds regardless of memory count.

MCP Server

Ferricula ships as an MCP server — drop it into Claude Code or any MCP-compatible agent as a persistent memory backend. Survives session restarts. Semantic search over past work. Agent identity that accumulates across days.