Agentic Memory

Agentic memory refers to the systems and architectures that enable AI agents to maintain persistent knowledge, learn from past interactions, and carry context across sessions and tasks. It's the capability that transforms an AI from a stateless query-response system into a persistent collaborator that accumulates understanding over time.

Current LLMs have a fundamental limitation: their context window is finite and ephemeral. A conversation ends, and the model retains nothing. Each new session starts from zero. Agentic memory addresses this gap through multiple complementary mechanisms.

Short-term working memory operates within a single task or conversation. The agent maintains a scratchpad of relevant information, intermediate results, and task state. This is analogous to the context window but structured and managed: the agent decides what information to keep, summarize, or discard as it works through complex problems. Techniques like ReAct (Reasoning + Acting) and chain-of-thought prompting create structured working memory within a single interaction.

Episodic memory stores records of past interactions, decisions, and outcomes. An agent with episodic memory can recall that a previous approach to a problem failed, that a user prefers a certain style of communication, or that a particular tool produced unreliable results. This memory is typically stored in vector databases using embeddings, enabling semantic retrieval of relevant past experiences when similar situations arise.

Semantic memory accumulates factual knowledge extracted from interactions and research. An agent working with a codebase builds a semantic memory of its architecture, conventions, and known issues. An agent assisting with business strategy accumulates understanding of the company's market position, competitors, and goals. This knowledge persists across sessions, allowing the agent to become more effective over time.

Procedural memory encodes learned skills and workflows. When an agent discovers an effective multi-step process for accomplishing a task, procedural memory allows it to replicate that process in future similar situations without re-deriving it from scratch. This connects to the concept of agent tool use and MCP Tool Use & Function Calling-based tool ecosystems.

The architecture of agentic memory intersects with RAG (Retrieval-Augmented Generation) but extends beyond it. Where RAG retrieves from a fixed knowledge base, agentic memory is dynamic: the agent writes to its own memory, reflects on stored experiences, and prunes or restructures its knowledge as understanding evolves. The generative agents research from Stanford demonstrated that agents with memory, reflection, and planning capabilities exhibit emergent social behaviors that agents without memory cannot.

Agentic memory is a key differentiator in the shift from AI tools to AI collaborators. An agent that remembers your preferences, understands your project's history, and learns from its own successes and failures provides compounding value over time — creating the kind of data flywheel that makes the agent-user relationship increasingly productive with each interaction.

Filesystem Memory

As of October 2026, the agent platforms increasingly expose memory as plain files that the agent reads and writes with its ordinary tools, alongside or instead of a vector store. In Anthropic's managed agents (memory is in beta), a memory store is a workspace-scoped collection of text documents mounted as a directory under /mnt/memory/ in the session's sandbox. Access is read_write by default or read_only, enforced at the filesystem level, and every change creates an immutable memory version, giving an audit trail and point-in-time recovery. The documentation caps a memory at 100 kB and a store at 10,000 memories, allows up to 8 stores per session, and recommends many small focused files. The OpenAI Agents SDK takes a similar route for sandbox agents: a MEMORY.md index, a short memory_summary.md injected at the start of a run, and per-run summaries kept in the workspace. Google's ADK instead offers a memory service the agent searches across past sessions. The file-based approach generalizes what CLAUDE.md does by hand, and it keeps memory inspectable: a person can open, diff and correct it.

Scheduled Consolidation

Memory written incrementally accumulates duplicates, contradictions and stale entries, so vendors have added a separate consolidation pass. Anthropic's "dreaming" (announced May 19, 2026; research preview) is described as a scheduled process that reviews sessions and memory stores, extracts patterns such as recurring mistakes and shared preferences, and curates memory. Per the documentation, a dream is an asynchronous job that takes one memory store and 1 to 100 past sessions and writes a new output store; the input is never modified, so the result can be reviewed and discarded. OpenAI's sandbox memory uses two phases, extraction of per-conversation notes and then a consolidation agent that runs when the sandbox session closes. Evidence of benefit is thin: Anthropic's announcement cites Harvey reporting roughly 6x higher completion rates in its tests, a customer figure relayed by the vendor rather than an independent measurement. Consolidation is distinct from context compaction, which summarizes history inside a single run, and it is one practical route to continual learning without changing model weights.

Provenance and Review

Writable memory is an attack surface and a source of quiet error. Anthropic's documentation warns that if an agent with read-write memory processes untrusted input, a successful prompt injection could write malicious content that later sessions read as trusted, and recommends read-only access for reference material. Consolidation adds a subtler risk. Zhan et al. (arXiv, August 2026) describe "authority collapse", where consolidation preserves a claim but erases the constraints on how its source may be used; on their AuthMem-Bench they observed it in 48 of 49 configurations (seven consolidators across seven model backbones), a mean unauthorized-action rate of 50.3% for collapsed memories without authority metadata, and a drop from 16.9% to 0.0% end to end when authority labels were predicted and persisted. This is a single preprint on a benchmark the authors built. The practical safeguards are consistent across sources: record which session or person wrote each memory, keep the link to its source, review consolidated output before it replaces the original, and, as OpenAI's documentation instructs its agents, treat memories as guidance and trust the current environment when the two disagree.