Overview

For agents to be genuinely useful, they need memory. Not just short-term context windows, but persistent, structured, and searchable memory that grows over time. This research explores architectures, retrieval strategies, and evaluation methods for long-term agent memory.

Key Questions

• What types of memory should agents have?
• How do we structure and retrieve memories effectively?
• How do we prevent forgetting while avoiding noise?
• How can memory be shared across agents safely?
• How do we evaluate memory quality and usefulness?

Current Hypotheses

• A hybrid approach (vector + graph + key-value) is optimal.
• Importance scoring > recency for long-term usefulness.
• Agents need consolidation, not just storage.

Approach

• Prototype a hybrid vector + graph + key-value store.
• Test importance scoring against simple recency baselines.
• Run agents on long-horizon tasks and measure recall accuracy.

Notes & Ideas

• Memory consolidation might borrow from spaced repetition.
• Could agents share a memory layer the way humans share institutional knowledge?
• Worth exploring forgetting curves as a deliberate compression mechanism.

Open Problems

• No standard benchmark for long-term agent memory yet.
• Balancing recall precision with retrieval latency at scale.
• Privacy and access control when memory is shared across agents.

References

• Park et al., “Generative Agents: Interactive Simulacra of Human Behavior” (2023)
• Packer et al., “MemGPT: Towards LLMs as Operating Systems” (2023)
• Zhong et al., “MemoryBank: Enhancing LLMs with Long-Term Memory” (2023)