Compare All 30 Memory Techniques

Single-table comparison across families. Skim columns to filter by your constraint, then click into the technique. Cross-linked from the main README.

A single table. Skim the columns to filter by your constraint, then click into the technique.

#	Technique	Family	Persistence	Retrieval	Token Cost vs. History	Best For
01	Conversation Buffer	Short-term	None	Append-only	Linear (grows forever)	Prototypes; short chats
02	Sliding Window	Short-term	None	Last k turns	Constant (fixed window)	Bounded chat budgets
03	Summary	Short-term	None	Compressed text blob	Sub-linear (summary cap)	Long chats where gist is enough
04	Summary Buffer	Short-term	None	Summary + recent	Sub-linear	Long chats where recent detail matters
05	Token Buffer	Short-term	None	Token-trimmed history	Constant (strict budget)	Hard token caps
06	Vector Store	Long-term	Disk / DB	Semantic similarity (Top-K)	Constant per call (K items)	Recall by meaning across many turns
07	Entity	Long-term	Disk (JSON)	Entity-name lookup	Tiny (per-entity record)	Track facts about people, places, things
08	Knowledge Graph	Long-term	Graph DB	Subgraph (1-2 hop)	Tiny (subgraph)	Multi-hop relational questions
09	Episodic	Long-term	Disk / DB	Time + semantic	Tiny (episode summary)	"What happened last week?"
10	Semantic	Long-term	Disk / DB	Semantic similarity over facts	Tiny (top facts)	Stable knowledge that survives sessions
11	Procedural	Long-term	Disk / DB	Task-type matching	Tiny (one procedure)	Skip re-deriving plans
12	Working Memory	Cognitive	None	Salience-ranked window	Bounded (priority queue)	Many simultaneous attention items
13	Hierarchical Layers	Cognitive	Multi-tier	Cache fall-through	Variable	Tame unbounded memory size
14	Consolidation	Cognitive	Disk / DB	Off-path background job	One LLM job per cycle	Reduce store noise over time
15	Compaction	Cognitive	Multi-level	Level-selected	Tiny if you pick L2/L3	Most queries do not need full detail
16	Self-Reflection	Cognitive	Disk / DB	Task-type retrieval	Tiny (1 reflection)	Improve from past failures
17	Memory Routing	Cognitive	Routes to others	Classifier-driven	Depends on target store	Decide which memory store to write/read
18	Temporal	Cognitive	Disk / DB	Time + semantic	Tiny	When facts have a "valid from / valid to"
19	Forgetting & Decay	Cognitive	Disk / DB	Strength-weighted	Shrinks over time	Bound long-running stores
20	Retrieval Patterns	Retrieval	Disk / DB	Hybrid (BM25 + dense + rerank + MMR)	Configurable	Quality retrieval at scale
21	Cross-Session	Retrieval	Disk / DB	Per-user isolated	Constant per user	Personal assistants
22	Multi-Agent Shared	Retrieval	Shared store	Permissioned	Constant	Several agents collaborate
23	Memory as Tools	Retrieval	Disk / DB	Agent invokes tools	Tool call latency	Agent decides when to read/write
24	Graphiti	Framework	Neo4j	Graph + temporal	Tiny	Production temporal knowledge graphs
25	Mem0	Framework	Mem0 cloud	Managed retrieval	API-priced	Drop-in personalization
26	Letta / MemGPT	Framework	Letta server	Three-tier (core/recall/archival)	Bounded	Long-running self-editing agents
27	Zep	Framework	Zep cloud / OSS	Temporal KG	API-priced	Conversational memory at scale
28	Evaluation	Production	None	Methodology	None	Measure quality before shipping
29	LoCoMo Benchmarks	Production	Benchmark fixtures	Benchmark scoring	None	Compare techniques on standard data
30	Production Patterns	Production	Tiered hot/warm/cold	Tier-routed	Cost-bounded	Ship to many users reliably

How to read the table. "Token Cost vs. History" describes how cost scales with the conversation length. "Tiny" means the cost is bounded by the number of items retrieved (K), not by total history. "None" means the row is methodology, not a runtime store.