Can AI Agents Remember What They Did Last Week

Why Agents Forget by Default

LLMs are stateless. When a session ends, the conversation context is discarded. The next session starts with only the system prompt and whatever static context the developer provides. There is no built-in mechanism for one session to pass knowledge to the next. This is a fundamental architectural property of how LLMs work: they process input tokens and produce output tokens, with no persistent state between invocations.

This means that an agent that spent an hour investigating a production incident on Tuesday starts completely fresh on Wednesday. It does not remember which services it checked, what anomalies it found, what hypotheses it tested, or what the resolution was. Every piece of context from Tuesday's investigation is gone unless it was explicitly saved somewhere outside the LLM's context window.

How Memory Enables Cross-Session Recall

A persistent memory system gives the agent a place to store knowledge that survives between sessions. The pattern is straightforward: during execution, the agent stores important observations, decisions, and outcomes to the memory system. At the start of each new session, the agent retrieves memories relevant to the current task. The retrieved memories are injected into the agent's context window, giving it access to knowledge from all previous sessions.

The quality of cross-session recall depends on three factors. First, what was stored: agents that store only high-value observations (facts, outcomes, surprises) produce better recall than agents that store everything (including intermediate reasoning and dead ends). Second, how it is retrieved: semantic search with cognitive scoring returns more relevant results than simple keyword matching. Third, how much context is available: the agent can typically load 10 to 50 past memories into its context without overwhelming the current task's context budget.

What Survives Between Sessions

With a well-configured memory system, agents can recall: facts about the environment (system configurations, dependencies, performance baselines), outcomes of past tasks (what worked, what failed, root causes of incidents), user preferences and patterns (communication style, priorities, common requests), entity relationships (which services connect to which databases, who owns what), and temporal events (what happened last Tuesday, the sequence of changes that led to an outage).

What does not survive, even with memory, is the reasoning process itself. The agent remembers that the root cause was a connection pool limit, but it does not remember the exact sequence of tool calls it used to discover that. This is by design: storing conclusions rather than reasoning produces a cleaner, more retrievable memory store. The agent can re-derive the reasoning if needed, and the conclusion is what matters for future tasks.

Adaptive Recall makes cross-session recall practical through cognitive scoring. When an agent asks "what did I find about Service X last week," the recall tool retrieves memories about Service X ranked by recency (last week's findings score higher than last month's), confidence (verified findings score higher than speculative ones), and entity connections (findings about Service X's dependencies also surface). The agent receives a focused, relevant set of past observations rather than a raw dump of everything it ever stored.