What is long-term memory in AI agents?
Long-term memory in an AI agent is information stored outside any single session, such as user preferences, past decisions or learned facts, that the agent writes deliberately and recalls in later sessions.
Short-term memory is the session: the history and state of one conversation or run, trimmed or summarised to fit the window. Long-term memory is what survives after that session closes, so the next one starts knowing something.
The labs split the two the same way. Google's ADK documentation calls session state short-term memory during one chat and the memory service a searchable long-term store. OpenAI's cookbooks treat trimming and summarisation as short-term tactics and a separate notes store as long-term memory. Anthropic's memory tool exists to carry learning across sessions.
Short-term and long-term memory
Short-term
One session's messages, tool calls and state. Managed by trimming, summarisation, compaction and context editing. Discarded or archived when the session ends.
Long-term
Selected facts promoted out of sessions into a durable store: files, a database, or a vector or keyword index. Recalled into later sessions when relevant.
The boundary
The hard design question is what crosses from one to the other. Storing whole transcripts is easy; storing the right facts, deduplicated and up to date, is the work.
Writing memories
Google's memory service ingests finished sessions; its managed Memory Bank uses a model to extract meaningful information and consolidate it with existing memories, while simpler services keep full conversations. OpenAI's personalization cookbook stages candidate notes during a session and consolidates them at the end, removing duplicates and letting the most recent update win conflicts.
The same cookbook sets write rules worth copying: store only what the user stated explicitly, never inferences, never instruction-like text, and never sensitive data such as payment details or account numbers.
Recalling memories
Reactive
The agent calls a memory tool when it recognises a gap, as with ADK's load memory tool or Anthropic's memory tool.
Proactive
The harness searches memory against the latest input and injects matches before the model runs, as with ADK's preload memory tool.
State injection
A compact profile and notes are rendered into the prompt at the start of every run, the approach OpenAI's personalization cookbook prefers when continuity matters more than breadth.
Where it fails
Stale facts recalled with confidence
The cookbook calls forgetting essential: without pruning, a store fills with superseded preferences. Recency weighting, confidence scores and expiry dates are the suggested remedies.
Over-influence
A remembered preference can override what the user is asking for now. The cookbook's precedence rule puts the latest user message above everything stored, and asks the agent to clarify on conflict.
Memory poisoning
Anything written to memory is replayed into future prompts, so instruction-like content stored once can steer every later session.
How Motiontic's engineers apply it
Motiontic's engineers treat long-term memory as a data product with an owner: a schema, write rules, retention limits, an audit trail and a way for users to see and delete what is stored. Memory has its own evals, covering what gets written, what gets recalled and whether recall changes the answer for the better.
Ways to recall long-term memory
| Reactive tool call | Proactive preload | State injection | |
|---|---|---|---|
| Who decides | The model | The harness, by search | The harness, always |
| Token cost | Only when used | Every turn, for matches | Every run, for the whole profile |
| Risk | Model forgets to look | Irrelevant matches crowd the prompt | Profile grows stale or too large |
| Fits | Large, rarely needed stores | Personal assistants and support | Small, high-value preference sets |
Common questions
What is the difference between short-term and long-term memory in AI agents?
Short-term memory is the history and state of one session, trimmed or summarised to fit the context window. Long-term memory is selected information kept after sessions end and recalled in later ones.
How do AI agents store long-term memory?
In files, databases or search indexes outside the model. Memories are written by the agent or extracted from finished sessions, then recalled by a tool call, a search before each turn, or injection at the start of a run.
How do agents forget outdated memories?
Through consolidation that removes duplicates and lets newer facts replace older ones, plus expiry dates, recency weighting and user-facing deletion. Without these, stale facts accumulate.
Related
Agent memory
Session state for the task at hand, plus stored knowledge the agent recalls later.
Memory tool
A file directory Claude can read and write across sessions, stored where the developer chooses.
Session (AI agents)
The durable log of a run, kept outside the window, that every prompt is built from.
Trimming vs summarization
Drop the old turns or compress them: the two basic short-term memory tactics.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- Google ADK docs, Memory
- Google, Architecting efficient context-aware multi-agent framework for production (2025-12-04)
- OpenAI cookbook, Context engineering for personalization: state management with long-term memory notes
- OpenAI cookbook, Short-term memory with sessions: trimming and summarization
- Claude blog, Managing context on the Claude Developer Platform (2025-09-29)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.