5-level lazy loading instead of flat RAG. On spawn, agents receive only L1 — one-line summaries of every memory entry.
When something looks relevant, they drill into L2, L3, etc. on demand. This keeps startup context small (~2k tokens
for 200 entries) while still having full verbatim detail available at L5.
Tab indentation maps directly to tree levels. No schema to think about — just write:
write_memory(prefix="L", content="Always restart MCP server after recompiling TS
Running process holds old dist — tool calls return stale results
Fix: kill $(pgrep -f mcp-server)")
Indented lines become children. Siblings at the same depth are siblings in the tree. append_memory adds children to
any existing node without overwriting.
Smart bulk-read selection. Not all L1 entries are shown equally. The V2 selection expands (shows children) for: newest
N entries per prefix, most-accessed M entries (time-weighted score), and all favorites. Everything else shows as a
compact title with a [+7 →] hint. This mirrors how human working memory works — recent and frequently-used things are
immediately accessible.
Obsolete, not deleted. When a lesson turns out wrong or a decision gets reversed, you mark it obsolete with a
correction reference ([E0076]). The old entry gets hidden from bulk reads but stays searchable. Past mistakes still
teach — you just don't pay tokens for them every session.
Per-agent isolation + optional shared store. Each agent has its own .hmem SQLite file. There's also a company.hmem
with role-based access control (worker/al/pl/ceo) for knowledge that should be shared across agents.
The whole thing is a single MCP server, ~3k lines of TypeScript, backed by SQLite with WAL mode. No vector embeddings,
no external services. Works offline.