Agent Memory Architectures

Memory systems transform LLM-based agents from stateless text generators into genuinely adaptive systems that persist and recall information across interactions. Du formalizes agent memory as a write-manage-read loop tightly coupled with perception and action — a framework that organizes the full design space through a three-dimensional taxonomy spanning temporal scope, representational substrate, and control policy.

Five mechanism families implement this loop:

1. Context-resident compression — Summarizing and compressing interaction history to fit within context windows. Progressive distillation of episodic memories into compact representations, trading information fidelity against context budget.

2. Retrieval-augmented stores — External memory (vector databases, knowledge graphs) accessed via embedding-based similarity search. Hybrid retrieval combines semantic similarity with temporal recency and importance scoring.

3. Reflective self-improvement — Agents generating insights and lessons learned from past experiences. Self-critique mechanisms identify errors and update behavioral policies. Meta-cognitive processes distill episodic memories into strategic knowledge.

4. Hierarchical virtual context — Multi-level hierarchies mimicking human short-term and long-term memory. Working memory for active reasoning, episodic memory for experiences, semantic memory for facts, with attention-based access across levels.

5. Policy-learned management — Learned policies for deciding what to store, when to consolidate, and how to retrieve. Reinforcement learning optimizes memory management strategies that evolve with agent experience.

Memory as a Versioned Platform Primitive (Jul 2026)

Agent memory crossed from research architecture into billable, versioned API surface during 2026. Anthropic's first-party Claude Platform changelog records a breaking beta-header migration on memory-store endpoints: agent-memory-2026-07-22 replaces managed-agents-2026-04-01 (sending both returns a 400), and it tightens listing semantics in ways that constrain how a memory store may be read:

• Results return in a stable server-defined order; order_by / order parameters are now ignored.
• depth accepts only 0, 1, or omission — other values are a 400.
• path_prefix must end in / and matches whole path segments, not substrings.
• Page cursors issued without the header are invalidated under it.
• All major SDKs (Python, TypeScript, Go, Java, Ruby, PHP, C#, CLI) send the new header by default; the prior behaviour retires 2026-07-22.

The design direction is legible from the constraints: hierarchical, path-addressed memory with shallow reads (depth 0 or 1) and segment-exact prefixes, ordered by the server rather than the caller. That is a store optimized for bounded, predictable retrieval cost — not for arbitrary client-side query. It is a commercial-infrastructure answer to the same write-manage-read problem the surveys formalize, and it constrains which of the five mechanism families a platform customer can actually implement on top.

Read alongside Agent Harness Evolution: memory is the harness component with the clearest productization path, shipping ahead of the evidence that automated harness design produces transferable gains.

This source is authoritative for what shipped, not for how well it works — it is procedural API documentation, not a benchmark disclosure.

A-MEM and Zettelkasten-Inspired Memory

A-MEM (Xu et al., NeurIPS 2025) offers a concrete implementation of agentic memory inspired by the Zettelkasten method. When new information arrives, the agent constructs a structured note with core content, contextual description, keywords/tags, and metadata. It then performs retrospective analysis — scanning existing memories for semantic connections and establishing bidirectional links. The system actively evolves: new information can modify contextual representations of existing memories, contradictory information triggers reconciliation, and frequently accessed memories gain higher retrieval priority.

Key Claims

• Write-manage-read loop is the fundamental memory abstraction — All agent memory systems can be understood as implementations of this three-phase cycle coupled with perception and action. Evidence: strong (Memory for Autonomous LLM Agents)
• Five mechanism families cover the design space — Context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, and policy-learned management. Evidence: strong (Memory for Autonomous LLM Agents)
• Zettelkasten-inspired memory outperforms fixed-structure baselines — A-MEM with dynamic note construction and linking beats flat, hierarchical, and summary-based memory on multi-session tasks across six foundation models. Evidence: strong (A-MEM: Agentic Memory for LLM Agents)
• Evaluation is shifting from static recall to multi-session agentic tests — Memory benchmarks now measure knowledge integration, temporal reasoning, and inference chains across sessions, not just retrieval accuracy. Evidence: strong (Memory for Autonomous LLM Agents)
• Agent memory now ships as a versioned, constrained platform primitive — Anthropic's agent-memory-2026-07-22 beta header replaced managed-agents-2026-04-01 on memory-store endpoints, fixing server-defined ordering, restricting depth to 0/1, and requiring segment-exact path_prefix matching. Evidence: strong for what shipped (first-party changelog); silent on efficacy (Anthropic Platform Release Notes)
• Sessions can override harness config without touching the base agent — Managed Agents sessions can override model, system prompt, tools, MCP servers, and skills per-session, making the memory-plus-scaffold bundle a runtime-configurable object. Evidence: strong (first-party changelog) (Anthropic Platform Release Notes)

Open Questions

• How to achieve continual consolidation without catastrophic forgetting?
• Can causally grounded retrieval replace similarity-based retrieval for better relevance?
• How to ensure self-generated reflective insights are reliable and grounded?
• How to extend memory systems to multimodal embodied settings (visual, spatial, proprioceptive)?
• How to secure memory stores against manipulation attacks (a documented agentic threat vector)?
• Do platform-imposed retrieval constraints (shallow depth, server-defined ordering, segment-exact prefixes) rule out whole mechanism families — notably policy-learned management, which needs control over retrieval order?
• Does per-session harness override turn production traffic into an implicit harness search, reintroducing the overfitting risk the harness-evolution audit identified — but with real users as the feedback signal?

Related Concepts

• Agentic Reasoning — Memory is the infrastructure for self-evolving agents (layer 2 of the three-layer framework)
• Reinforcement Learning for Agents — RL optimizes memory management policies
• Agent Safety & Alignment — Memory manipulation is a documented attack surface
• Agent Harness Evolution — memory is the harness component furthest along the productization path

Backlinks

Pages that reference this concept:

• Agentic Reasoning
• Agent Safety & Alignment
• Agent Harness Evolution

Changelog

• 2026-07-22 — Compiled new source (anthropic-claude-platform-updates-2026-07): added "Memory as a Versioned Platform Primitive" section documenting the agent-memory-2026-07-22 beta-header migration and per-session harness override; two new claims and two open questions on platform-imposed retrieval constraints