Letta / MemGPT
LLM-as-Operating-System paradigm — model manages its own memory, paging between core context, recall buffer, and archival store. UC Berkeley origin; transitioning to letta_v1_agent .
At a glance
- Type
- Tiered: core / recall / archival
- Tier
- T1
- Section
- Dedicated memory layers
- Created
- 2023-10
- Latest release
- 0.16.7 2026-03-31
- License
- Apache-2.0
- GitHub
- 22.5k★ +151/mo Python
- Pricing
- Free + paid
- Funding
- $10M total $70M val Seed · 2024-09
Taxonomy
- storage
- kv
- retrieval
- similarity
- persistence
- long-term
- update
- agent-controlled
- unit
- episode
- governance
- inspectable
- conflict
- llm-arbitrate
When to use
Optimised for: LLM-as-OS paradigm; agent autonomy in managing its own memory
Anti-fit: no anti-fit explicitly stated
Pros & cons
Pros
Most architecturally principled — explicit OS-style tiering makes capacity behavior predictable; UC Berkeley origin + active v1 maturity.
Cons
Tier configuration adds operational complexity vs flat vector memory; latency is sensitive to paging decisions on long histories.
Claims & capabilities
21.8k★. $10M seed at $70M (Sep 2024). Reports ~83.2% on LongMemEval; ~74.0% on LoCoMo with gpt-4o-mini using files-as-memory.
Technical surface
- API surface
- REST, SDK: Python, TS
- Backend storage
- Postgres + pgvector
- Deployment
- Both
- Embedding model
- multiple supported
- Multi-tenancy
- Logical namespace per org/project/agent; self-hosted Docker option for hard tenant isolation
- MCP
- native (first-party) — Letta serves & consumes MCP
- A2A
- not documented publicly
- OpenTelemetry
- not documented publicly
Similar systems
Other dedicated memory layers in the catalog, ranked by inbound references.
- Mem0 T1
Universal memory layer for AI agents. Three concurrent stores (vector + graph + KV); LLM-extracted facts; concurrent retrieval via ThreadPoolExecutor.
- Zep & Graphiti T1
Bi-temporal knowledge graph (event time + ingestion time). Strong on chronological reasoning and contradiction tracking. Graphiti is the open-source core.
- Cognee T1
"Extract–Cognify–Load" pipeline that turns raw input into a typed, queryable knowledge graph for agent recall.
- Hindsight (Vectorize) T1
Standalone memory service from Vectorize. Open source. Biomimetic four-network design (World, Bank, Observation, Opinion). Ships an MCP memory server.
- Memvid T2
Single-file memory layer (one .mv2 file). No DB, no server. Append-only sequence of immutable Smart Frames with timestamps + checksums. Native Rust core (rewritten from Python).
- Supermemory T1
Memory engine with API, app, browser extension, and MCP server. Extracts facts, tracks updates, resolves contradictions, auto-forgets expired info. Plugins for Claude Code, OpenCode, OpenClaw, Hermes.
Related systems
References (3)
- Anthropic Claude (foundation models) depends on at runtime — adjacent-infrastructure cell: BYO LLM (OpenAI/Anthropic/local); bundled storage (Postgres/SQLite)
- OpenAI GPT family (GPT-5 / GPT-4o / o3 / o4) depends on at runtime — adjacent-infrastructure cell: BYO LLM (OpenAI/Anthropic/local); bundled storage (Postgres/SQLite)
- pgvector depends on at runtime — backend-storage cell: Postgres + pgvector