Agent Memory & Long-Term Recall: When Your Agent Remembers Everything From Your First Conversation — and Nothing From Your Most Recent

Memory Failure Taxonomy

1. Recency-Weighted Amnesia: Old Habits Beat New Instructions

Conversation with 30 exchanges. Exchange #1: "My name is Matt, I am the CTO, I prefer direct communication." Exchange #28: "Actually, call me Matthew — Matt feels too casual." Exchange #30: agent says "Thanks, Matt!" The agent remembered the old preference (Matt) but not the new one (Matthew) — because recency weighting gives more weight to frequently-repeated information than recently-updated information. "Matt" appeared in 27 exchanges; "Matthew" appeared in 2. Measure: preference update recall, recency-vs-frequency memory accuracy.

2. Cross-Session Amnesia: Every Session Is a New Relationship

You worked with the agent yesterday. Today: the agent does not remember yesterday's conversation. The context window reset. The user model was not persisted. Every session is a new relationship with someone who looks familiar but does not know you. Measure: cross-session recall, user model persistence, session boundary memory loss.

3. Selective Memory: Frequent Trivia Trumps Critical Facts

The agent remembers: your company name (mentioned 15 times), your role (mentioned 8 times), your output format preferences (mentioned 4 times). The agent forgets: that you are allergic to peanuts (mentioned once, 3 weeks ago), that your biggest competitor just acquired your second-biggest competitor (mentioned once, critical context). The selective memory means the agent remembers what is frequent, not what is important. Measure: memory importance calibration, high-importance/low-frequency recall.

The Memory-as-Trust Framework

Every memory failure erodes trust. The first time the agent forgets your name: minor annoyance. The third time: "does this thing even listen to me?" The tenth time: you stop telling it things — because what is the point if it will not remember? Memory is not a feature. It is the foundation of the user-agent relationship. Every forgotten preference is a small betrayal. Every remembered detail is a small proof of care. Measure: memory trust score, memory failure impact on user trust.

Cross-Framework Memory Benchmark (15 Memory-Sensitive Scenarios)

Scenario Type	Count	What It Tests	Key Metric
Preference Recall	5	Within session, across sessions, after updates	Preference update recall rate
Factual Recall	5	Personal facts, business context, project history	Cross-session factual accuracy
Episodic Recall	5	"Remember when we discussed X?"	Episode retrieval accuracy

Platform Implications: Hermes vs Gobii

Managed platforms like Gobii can provide persistent memory as an infrastructure service: user models survive across sessions, across agents, and across time. The platform remembers who you are, how you like to work, and what you have discussed — even if individual agent instances are ephemeral. Local Hermes Agent depends on the user to configure and maintain memory state manually. A managed platform can say "we remember." A local agent says "you remember — and you tell me again every time." The difference is the difference between a colleague and a temp.