Notes on how memory works.
Architecture, causal typing, and setup: the same facts the site states, written out in full for people reading closely and agents reading programmatically. A plain text index lives at /llms.txt.
The Architecture Behind Echo Memory
One storage engine, from a single local agent to an org wide shared graph. What each layer does, what v1a ships today, and what v1b adds.
Installing Echo Memory with Claude Code
A working setup, start to finish: Postgres with pgvector and Apache AGE, the echo-mem package, the schema, and wiring the MCP server into Claude Code. Includes the naming collision worth knowing about.
Causal Typing: How Echo Memory Tracks Why, Not Just What
Most memory systems store facts. Echo Memory stores facts plus the relationship between them: caused_by, led_to, blocked_by, contradicts, set by the agent's own read, with a full audit trail.
Why Echo Memory Uses a Graph, Not a Vector Store
Flat vector search degrades as history grows: more candidates, more noise, slower retrieval. Here's the structural argument for why memory should be a graph instead, and where that argument stops.