Agent Memory

Definition

Agent memory persists and retrieves task-relevant state beyond the immediate model call. It may hold working state within a run, episodic history across interactions, procedural knowledge, or user-specific facts; the defining behavior is durable recall, not merely a long prompt. LinkedIn explicitly separates working, episodic, procedural, experiential, and collective memory. LinkedIn

Corpus Support

The analysis counts 183 deduplicated agentic systems (18.5%) across 15 industries. Evidence consists of 115 article, 141 transcript, and 24 summary quotes; 279 are stated and 1 implied. Corpus analysis

Problem It Solves

Stateless model calls lose conversation progress, prior tool results, user preferences, and learned procedures. Block stores session and message history around durable workflows, while Prosus bootstraps useful personalization from existing product history instead of waiting for a new agent to accumulate memories. Block Prosus

Applicability Signals

• Work spans multiple turns, retries, or asynchronous pauses. Block
• The system should recall user or organization preferences across sessions. Harvey
• Agents need reusable procedural knowledge or prior outcomes. LinkedIn
• Cold-start quality can benefit from pre-existing product history. Prosus

Architecture and Variants

• Session memory: persist messages and intermediate state keyed by task or conversation. Block
• Layered memory: separate short-lived working state from episodic, procedural, and collective stores. LinkedIn
• Profile memory: retrieve preferences and prior interactions to personalize a new request. Harvey
• Platform-managed state: use an agent runtime to store session, conversation, and intermediate workflow state. Iberdrola

Why Teams Choose It

Memory makes long-running and personalized behavior possible without replaying all historical data into every prompt. It also lets products reuse knowledge already captured elsewhere, reducing the cold-start burden for a new agent. Prosus Harvey

Tradeoffs

• Persistent memory improves continuity but creates privacy, retention, and deletion obligations. Harvey
• Storing more history increases recall opportunities but can retrieve stale or irrelevant state; layered memory narrows retrieval scope. LinkedIn
• Runtime-managed memory simplifies lifecycle handling but couples state semantics to a platform. Iberdrola

Gotchas and Resolutions

• Conversation persistence is not automatically durable user memory; Manus explicitly lacked persistent cross-conversation memory despite session continuity. Manus
• Replaying raw history can bloat context; retrieve only the memory needed for the current decision. LinkedIn
• Workflow state and model-visible memory have different consistency needs; Block separates durable execution state from the context supplied to inference. Block

When Not to Use It

Do not persist memory for one-shot tasks, or when sensitive information cannot be retained with a clear user benefit. Manus demonstrates a valid intermediate design: preserve task continuity while declining to claim cross-conversation memory. Manus

Boundaries and Related Patterns

context-engineering governs what information is assembled for a model call; agent-memory is a persistent source that context engineering may query. context-compaction reduces active context size and may summarize memory, but it does not define retention or recall. durable-execution preserves workflow progress and retries; not all durable state should become model-visible memory. Block

Operational Checklist

• Define memory scope, owner, lifetime, and deletion behavior. Harvey
• Separate workflow state, conversation history, user profile, and learned procedure stores. LinkedIn
• Retrieve narrowly and record why each memory entered context. Prosus
• Test stale, conflicting, missing, and unauthorized memories. Block

Representative Implementations

Methodology and Limitations

This page samples 183 deduplicated systems. Published descriptions often use “memory” loosely, so the boundary emphasizes persistence and retrieval; support frequency does not establish that longer retention improves outcomes. Corpus methodology