Research Behind Agentverse Memory

Long-term memory for LLM agents requires systems that remain precise at million-token scale — a challenge no existing graph-based system has publicly met. We present PheromGraph, a persistent agent memory system that applies stigmergic pheromone dynamics to a knowledge graph, enabling retrieval quality to evolve through use. PheromGraph deposits five typed pheromone traces (episodic, semantic, procedural, working, meta) on nodes and edges during every retrieval, uses exponential decay with type-specific time constants τ, and traverses the graph using A* with a composite cost function combining semantic similarity, pheromone weight, and structural proximity.

• ▸First formalization of stigmergic pheromone dynamics in LLM agent knowledge graphs — no prior art in production systems
• ▸5-type pheromone taxonomy with type-differentiated exponential decay (τ per memory type)
• ▸A* semantic pathfinding with pheromone-weighted composite cost: semantic (40%) + pheromone (30%) + structural (30%)
• ▸Closest prior work (arXiv:2512.10166, Dec 2024) uses grid simulations only — no LLMs, no semantics, no knowledge graphs
• ▸Primary benchmark: BEAM (first graph-system to publish results) — target >75% BEAM 1M, >64.1% BEAM 10M (vs Hindsight SOTA 73.9% / 64.1%)
• ▸Supporting: LongMemEval_S target ≥94%; LME-V2 LAFS metric rewards our <20ms retrieval directly
• ▸Code to be open-sourced on submission: github.com/fetchai/agentverse-memory

We reformulate agent memory as an active inference problem under the Free Energy Principle: storing a memory minimizes surprise about future queries; retrieving a memory minimizes surprise about the current context. This framing unifies episodic, semantic, procedural, and working memory under a single variational objective, explains pheromone dynamics as a precision-weighting mechanism, and yields novel predictions about optimal memory consolidation schedules.

• ▸Free Energy Principle (Friston 2010) applied to LLM agent memory — novel theoretical framing
• ▸Unifies all 4 memory types under a single variational objective
• ▸Explains pheromone dynamics as precision-weighting in predictive coding terms
• ▸Companion paper to PheromGraph — shares experimental infrastructure

No benchmark currently measures the correctness and efficiency of shared memory between cooperating LLM agents. MAS-MemEval fills this gap: 500 multi-agent scenarios across 5 task types (collaborative research, orchestrator-worker task delegation, adversarial privacy, temporal consistency across agents, and knowledge synthesis). We establish baselines for Agentverse Memory shared spaces, vector-only systems, and no-shared-memory ablations.

• ▸First benchmark for multi-agent shared memory — gap confirmed, no prior work exists
• ▸500 scenarios × 5 task types = 2,500 evaluation instances
• ▸Includes adversarial privacy scenarios (agent A should not access agent B's private memories)
• ▸Temporal consistency task: does shared memory stay coherent when two agents update the same fact concurrently?
• ▸Agentverse Memory is the reference implementation (shared spaces feature, Builder plan+)