Hindsight: Agent Memory with Retain, Recall, Reflect
Source: https://arxiv.org/abs/2512.12818 Author: Chris Latimer, Nicoló Boschi, Andrew Neeser, Chris Bartholomew, Gaurav Srivastava, Xuan Wang, Naren Ramakrishnan Date: 2025
Summary
Novel agent memory architecture organizing memory into four logical networks (world facts, agent experiences, entity summaries, evolving beliefs) with three operations: retain, recall, reflect. Achieves 83.6% on LongMemEval with a 20B model, vs 39% baseline and matching full-context GPT-4o.
Key Claims
- Four memory networks: world facts, agent experiences, entity summaries, evolving beliefs
- Three operations: retain (add), recall (access), reflect (update/synthesize)
- LongMemEval: 83.6% (20B model) vs 39% baseline — matching full-context GPT-4o
- 91.4% on LongMemEval and 89.61% on LoCoMo with larger backbone
- Enables agents to explain reasoning coherently across extended sessions
- Goes beyond simple snippet extraction — structured, queryable memory bank
Connection to Other Sources
Key empirical data point for Agent Memory. Compare to Mastra observational memory (automatic extraction) and RL-trained memory agent.
Concepts
- Agent Memory — four-network structured memory architecture
- Context Engineering — memory as pre-organized context