We introduce a hybrid vector-symbolic memory architecture that enables persistent episodic memory in production AI agents. Our approach combines dense retrieval with symbolic consolidation, achieving 3.2× better recall over 30-day horizons while reducing memory footprint by 41%. Evaluated on six enterprise agent benchmarks.
The challenge of maintaining coherent long-term memory in AI agents represents one of the fundamental limitations of current large language model deployments. While transformer-based models excel at in-context reasoning, their fixed context windows prevent them from accumulating experience across sessions — a property essential for agents operating in persistent enterprise environments. Prior work has addressed this through external memory stores, but naive approaches suffer from retrieval precision problems at scale, semantic drift over time, and prohibitive storage costs. In this paper, we introduce a hybrid architecture that combines dense vector retrieval with symbolic knowledge consolidation, enabling agents to maintain accurate, queryable episodic memory over multi-month operational horizons.
Our system comprises three core components: a working memory buffer that handles within-session context, an episodic store built on HNSW vector indices with metadata filtering, and a semantic consolidation layer that periodically abstracts and compresses stored episodes into symbolic representations. The consolidation process runs asynchronously, triggered when episodic store utilization exceeds a configurable threshold. During consolidation, a fine-tuned compression model summarizes clusters of semantically related episodes into higher-level "memory traces" — preserving causal relationships and factual content while discarding redundant detail. These traces are indexed alongside raw episodes and participate in retrieval through a learned fusion scoring function.
We evaluate our architecture on six enterprise agent benchmarks spanning customer support, research assistance, project management, and code review scenarios. Each benchmark consists of 90-day agent operation logs with human-annotated ground truth for memory-dependent queries. Our approach achieves 3.2× better recall at 30-day horizons compared to naive vector retrieval baselines, with 41% reduction in memory footprint attributable to consolidation. Latency overhead at query time is 12ms (p99) on standard inference hardware, acceptable for all evaluated use cases. Ablation studies confirm that both the consolidation layer and the fusion scoring function contribute meaningfully to end-to-end performance.
Table 1 summarizes task completion rates across all six benchmarks. Our method achieves an average improvement of 28 percentage points over the no-memory baseline, with the largest gains in scenarios requiring multi-session context (customer support: +41pp, research assistance: +37pp). Importantly, precision remains stable over the evaluation horizon — a key failure mode in baseline systems where retrieval quality degrades as stores grow. We attribute this to the consolidation mechanism's ability to reduce index density while preserving semantic coverage.
We have presented a hybrid memory architecture that addresses the core limitations of existing agent memory systems: precision degradation at scale, high storage costs, and poor long-horizon recall. Our approach is model-agnostic, operates within standard inference infrastructure, and is available as an open-source library (agent-memory on PyPI). Future work will explore personalized consolidation strategies, cross-agent memory sharing, and memory-grounded chain-of-thought reasoning. We believe persistent episodic memory is a prerequisite for the next generation of enterprise AI agents, and this work takes a concrete step toward making it production-ready.