agentmemory-architecture
agentmemory is a memory server for coding agents. It runs locally, captures observations, indexes them for hybrid retrieval, and serves them back over REST and MCP. Worker Compose runs the pinned iii engine with AgentMemory and the registry state and http workers. The state worker uses a file-backed KV store by default, with Redis available by configuration; it is not SQLite.
iii primitives
Everything is a function, a trigger, or worker state on the iii engine. There is no separate plugin system; the worker registers functions (mem::*) and HTTP triggers (api::*) and the engine routes calls. agentmemory does not bypass iii; new capability is a new function plus a trigger.
Retrieval model
Recall combines BM25 keyword search with configured vector similarity and graph expansion. A clean installation has no embedding provider and uses BM25 without an API key. Set EMBEDDING_PROVIDER=local to enable on-device embeddings, which download a model once. A remote embedding provider runs only when EMBEDDING_PROVIDER names it. LLM summaries, consolidation and reflection run only when AGENTMEMORY_LLM_PROVIDER selects a provider or AGENTMEMORY_ALLOW_AGENT_SDK=true turns on the Claude Agent SDK fallback; a key in the environment alone does not turn them on. Per-observation LLM compression also needs AGENTMEMORY_AUTO_COMPRESS=true.
Data boundary
Captured prompts, tool inputs and outputs, and memories stay in the local store. They leave the machine only through options the user turns on: a selected LLM provider, the Agent SDK fallback (AGENTMEMORY_ALLOW_AGENT_SDK=true), a remote embedding provider, a Redis state backend, or mesh sync to peer servers the user registers.
Storage and lifecycle
Memories carry content, concepts, files, importance, and timestamps, grouped into sessions and optionally linked to commits. A lifecycle of capture, compress, consolidate, and forget keeps the store useful over time rather than letting it grow unbounded.