Source: Letta documentation, memory blocks and archival memory
Letta's docs on in-context memory blocks (label, description, value, limit, read_only, shared), searchable archival memory and conversation search.
Contents
- What: official framework documentation (Letta is the company and framework that grew out of MemGPT)
- Publisher: Letta
- Date: undated (read 2026-11-14)
- Original: Memory overview · Memory blocks · Archival memory
Summary
Letta organises agent memory as editable blocks pinned into the context window, plus an out-of-context archival store the agent searches with tools, plus automatic search over past messages. All state (messages, memories, tool calls) is persisted in a database, so evicted messages remain retrievable.
Key claims
- A block has a
label, adescription(what the agent relies on most to decide how to use it), avalue, a characterlimit, and an optionalread_onlyflag. - Attached blocks are prepended to the prompt in a
<memory_blocks>section showing description, current/limit character counts and value; they never need retrieval. - A block can be attached to several agents; one update is visible to all of them (shared memory). Setting
valuereplaces the whole block, so concurrent writes are last-write-wins. The docs give no defaultlimit. - Archival memory is semantically searchable, effectively unlimited, taggable, and cannot be pinned in context. Tools:
archival_memory_insertandarchival_memory_search(tag filter, pagination). - Guidance: blocks for always-visible, frequently changing information; archival for documents, logs, reports and history. Archival is curated by the agent; conversation search retrieves what was said automatically.
- Letta’s newer docs describe “dreaming”: background subagents that review recent conversations and consolidate memory without blocking the active agent. Triggers are off, every N user messages, or on context compaction. Dreaming in Letta Code is distinct from the older, experimental server-side sleep-time agents.