Memory Types · Comparison
Short-Term vs Long-Term Memory in AI Agents
Short-term memory holds the current conversation in the context window; long-term memory persists facts and history across sessions — agents need both, plus consolidation to move information between tiers.
Two-tier model
Model
The two-tier memory model
Tier 1: working/short-term (hot, volatile). Tier 2: long-term (cold, durable). Information flows via consolidation.
Analogous to OS paging (MemGPT) and human memory metaphors — but technically: context window = short-term; vector/graph store = long-term. See memory hierarchy for the full stack.
Comparison
Short-term vs long-term: comparison table
Short-term = now; long-term = tomorrow.
| Dimension | Short-term | Long-term |
|---|---|---|
| Duration | Current session / turn | Days to years |
| Location | Context window, Redis buffer | Vector DB, graph, memory API |
| Capacity | 128K–1M tokens (model-dependent) | Effectively unbounded |
| Update rate | Every message | On consolidation / write |
| Backend | LLM context, Redis lists | Engram, Mem0, Zep, pgvector |
| Example content | “User just said they’re vegetarian” | “User prefers email, vegetarian since 2024” |
Consolidation
How information moves from short-term to long-term
Triggers: session end, salience threshold, explicit memory tool. Steps: extract → embed → store → optionally summarize source.
- Detect salient facts in current context
- Extract via LLM or framework pipeline
- Embed and write to LTM store
- Optionally summarize or evict raw transcript from short-term
Mem0 uses ~1,800 vs ~26,000 tokens per LOCOMO query after selective LTM retrieval (Chhikara et al., 2025).
Production
When agents need both tiers
Stateful agents always need both: context for reasoning now, LTM for tomorrow.
Single-shot Q&A may skip LTM. Any multi-session assistant, support bot or personalization system needs short-term context + durable long-term store. See why AI agents need memory.
Frameworks
Framework support for two-tier memory
- Engram — Weaviate LTM vectors + your context window as STM
- Letta — explicit core (STM) + archival (LTM) tiers with paging
- Mem0 — LTM API + conversation context
- Zep — temporal graph LTM + session context
- LangMem — checkpointer (STM) + store (LTM)
FAQ
Frequently asked questions
What is the difference between short-term and long-term AI memory?
Short-term = current conversation in the context window (volatile). Long-term = persisted facts in external storage across sessions (durable). Agents need both plus consolidation between tiers.
Is AI short/long-term memory the same as human memory?
Inspired by the metaphor but technically different — AI short-term is the LLM context window; long-term is vector/graph stores. No biological equivalence.
Is the context window the same as short-term memory?
Yes — the context window is the primary short-term memory substrate. Redis buffers and working memory extend it but context is the core STM layer.
When should agents consolidate short-term to long-term?
Session end, salience thresholds, memory pressure or explicit memory-tool calls. See memory consolidation.
Can agents skip long-term memory?
Single-shot Q&A can. Any multi-session assistant needs LTM — otherwise the agent forgets everything between sessions.
Does Mem0 support both memory tiers?
Yes — Engram and Mem0 both provide LTM extract/store/retrieve; you pair with conversation context as STM. Engram is Weaviate-native; Mem0 is framework-agnostic.
How does Letta handle memory tiers?
Explicit core memory (STM, in-context) and archival memory (LTM, paged in/out). MemGPT DMR 93.4% (Packer et al., 2023). See Letta alternatives.
Does Claude have short and long-term memory tiers?
Claude's context window acts as STM. Long-term memory requires external tools (Engram, Mem0, etc.) — not built into the model API alone.
Is LSTM the same as AI agent long-term memory?
No — LSTM is a recurrent neural network architecture. Agent LTM refers to external vector/graph stores, not RNN hidden states.
How do you benchmark both memory tiers?
LOCOMO and LongMemEval test multi-session recall (LTM). Measure token usage for STM vs full-context. Mem0 ~1,800 vs ~26,000 tokens (Chhikara et al., 2025).