AI Agents Memory

A hands-on lab about memory for AI agents: what to store, where to store it, and how to retrieve it. The diagram below maps agent memory into six technical levels, from the active context window up to cross-session reflection and self-editing policies.

High-level overview: https://mem.codedorf.dev

The six levels

[ LEVEL 6: METACOGNITIVE / EVOLUTIONARY LAYER ]
  │  Scope: Autonomous system-level adaptation (Cross-Session / Global)
  │  Medium: Offline async batch jobs, reflection graphs, self-editing system instructions
  │  Function: Synthesizes high-level principles from experiences; refines tool policies
  │
  ▼ updates & prunes
[ LEVEL 5: PROCEDURAL / SKILL MEMORY ]
  │  Scope: Behavioral patterns, execution workflows, tool schemas (Persistent)
  │  Medium: System prompt registries, MCP tool catalogs, code execution sandboxes
  │  Function: Encodes "how to do things" (few-shot patterns, task execution graphs)
  │
  ▼ grounds & validates
[ LEVEL 4: SEMANTIC / DECLARATIVE MEMORY ]
  │  Scope: Universal world knowledge, user entities, domain facts (Persistent)
  │  Medium: Knowledge graphs (triplets), relational DBs, dense vector embeddings
  │  Function: Encodes "what is true" independent of temporal sequence (e.g., user constraints)
  │
  ▼ consolidates & indexes
[ LEVEL 3: EPISODIC / AUTOBIOGRAPHICAL MEMORY ]
  │  Scope: Historical narrative, run traces, chronological interactions (Time-Anchored)
  │  Medium: Append-only event logs, time-decayed vector stores, execution histories
  │  Function: Encodes "what happened and when" (past failures, session logs, tool I/O)
  │
  ▼ retrieves & compacts
[ LEVEL 2: SHORT-TERM / CONVERSATIONAL BUFFER ]
  │  Scope: Active thread or multi-step task run (Session Scope)
  │  Medium: Sliding message array, thread caches, recursive rolling summaries
  │  Function: Tracks immediate dialogue continuity and the active task boundary
  │
  ▼ loads into
[ LEVEL 1: WORKING MEMORY / SCRATCHPAD ]
     Scope: Current forward pass / single step execution (Immediate)
     Medium: Active Context Window (Tokens, Attention, KV Cache)
     Function: Holds chain-of-thought, current tool inputs/outputs, and active sub-goals

Comparison table

How each system maps to the six levels. Deep dives live in mds/.

Legend: ● first-class, ◐ partial or possible with work, - not built in.

System	Type	L1	L2	L3	L4	L5	L6	Storage and retrieval	Best for
Memanto	Memory agent	-	-	◐	●	-	◐	Local vector store, typed facts	Low-ops factual memory
Mem0	Memory layer	-	◐	◐	●	-	◐	Vector plus optional graph	Drop-in personalization
Zep	Temporal KG platform	-	◐	●	●	-	-	Bi-temporal knowledge graph	Facts that change over time
Letta	Agent runtime	●	●	●	●	◐	◐	Core blocks, archival, database	Self-editing long-lived agents
MemOS	Memory OS	●	●	●	●	●	◐	MemCube over vector, graph, KV, LoRA	Fleet-scale memory infra
TencentDB Agent Memory	Team memory hub	●	●	●	●	●	◐	SQLite plus sqlite-vec and files	Local and team agents
LangMem	SDK	◐	◐	●	●	●	◐	Bring your own store	LangGraph apps
Cognee	Graph memory engine	-	-	◐	●	-	◐	Graph plus vector plus relational	Document and enterprise KG
Graphiti	Temporal KG library	-	-	●	●	-	-	Neo4j, FalkorDB, Kuzu, Neptune	Custom temporal graph
A-MEM	Research system	-	-	●	●	-	◐	ChromaDB plus linked notes	Adaptive note networks
MemoryOS	Research system	●	●	●	●	-	◐	Tiered local store	Personalized long dialogs
HippoRAG	Retrieval framework	-	-	◐	●	-	-	Graph plus Personalized PageRank	Multi-hop document QA
Memobase	Profile backend	-	◐	●	●	-	-	Postgres plus Redis	Consumer chat profiles
Supermemory	Memory API	-	◐	●	●	-	◐	Graph plus vector, one store	Product memory and RAG
MemMachine	Memory layer	●	●	●	●	-	-	Client-server episodes and profile	Ground-truth recall

Licenses, hosting, and papers:

System	License	Hosting	Paper
Memanto	MIT	Local	arXiv:2604.22085
Mem0	Apache-2.0	Cloud or self-host	arXiv:2504.19413
Zep	Apache-2.0 (Graphiti)	Cloud, engine self-host	arXiv:2501.13956
Letta	Apache-2.0	Cloud or self-host	MemGPT, arXiv:2310.08560
MemOS	Apache-2.0	Cloud or self-host	arXiv:2507.03724
TencentDB Agent Memory	MIT	Local	none
LangMem	MIT	Library	none
Cognee	Apache-2.0	Cloud or self-host	none (used in arXiv:2505.24478)
Graphiti	Apache-2.0	Self-host	Zep paper
A-MEM	MIT	Local	arXiv:2502.12110
MemoryOS	Apache-2.0	Local	arXiv:2506.06326
HippoRAG	MIT	Library	arXiv:2405.14831
Memobase	Apache-2.0	Cloud or self-host	none
Supermemory	MIT	Cloud or self-host	none
MemMachine	Apache-2.0	Cloud or self-host	arXiv:2604.04853

Benchmarks are labeled in each deep dive as paper results or vendor claims. Harnesses differ, so do not compare scores across papers without checking the setup.

What's next

• mds/ - technical deep dives on individual memory systems
• lab/ - the hands-on lab: does memory make the second run better?
• INTENT.md - project intent and roadmap
• docs/index.html - high-level overview published on GitHub Pages