mi∃a

miea — graph memory for AI agents. You ∃ in your agent's memory.

Structure and pointers only — all interpretation happens in the consuming agent's LLM at read time. No pre-computed summaries; embeddings only as an optional derived cache. Design notes: DESIGN.md (local).

Model

• Nodes = nouns, named content records
• Edges = verbs (persists_with, contradicts, …) — traversal paths concatenate into sentences: [Postgres] --persists_with--> [WAL]
• Nested graphs: containment hierarchy; every node at any depth is a full memory, not a folder
• Signposts: parents list their children as ranked references (label + breadth score), never generated summaries

Quick start

uv tool install --with sentence-transformers .   # from this repo
miea setup                                       # wizard: path + name → MCP JSON

Or manually:

miea init MyMemory --root ~/Documents/my_memory
miea --root ~/Documents/my_memory add Postgres --content "relational database"
miea --root ~/Documents/my_memory link Postgres persists_with WAL
miea --root ~/Documents/my_memory land Postgres      # rideable payload + signpost
miea --root ~/Documents/my_memory steer Postgres WAL # one branch of the slide

Architecture

JSON files (durable truth) → in-memory graph/indexes → agent via CLI/MCP

Files are never grepped or dumped raw — the core returns shaped payloads only. Ranking (breadth × recency) orders results but never filters: all memories are equal in importance; low rank costs latency, never visibility.

Status

Core complete (47 tests): model, store, waterslide read loop with signpost paging, write tier with guardrails, LCA, promotion-split, provenance audit, epistemic verify pass, hybrid FTS+vector search, wakeup snapshot, setup wizard, CLI + MCP server with read/write tiers.