neuron-mr-white/unipi/memory

Memory

Persistent cross-session memory backed by a shared MemPalace palace. The markdown files under ~/.unipi/memory/ are the durable tier; the palace is the searchable tier. Drawers written by other agents (Devin, zcode, diaries) are searchable alongside pi's own — every result carries its source.

Files

Each memory is one markdown file:

~/.unipi/memory/<project>/<type>/<id>.md

with YAML frontmatter (id, title, tags, project, created, updated, type) where type ∈ preference | decision | pattern | summary, and a per-project mempalace.yaml (wing + the 4 type rooms). project = sanitized basename of the session's cwd — the same rule other MemPalace agents use.

Tools

memory_search covers all drawers — results are labeled wing › room · source (pi, devin, zcode, mempalace, …) and chunks of the same file are grouped into one hit. Use scope: "project" to restrict to the current project's wing.

Store outcomes

memory_store always writes the markdown file. The palace write is reported honestly:

• filed ✓ — the daemon mine job finished and the drawer exists
• queued ⧗ — the job was accepted but didn't finish in time (it still lands)
• markdown only ⚠ — no daemon reachable; journaled to .pending.json and

replayed when a daemon appears

Switches (/unipi:settings → Memory)

~/.config/mempalace/agent-hooks.json {enabled:false} mutes both reminders. /unipi:memory status shows backend/daemon/reader/counts/pending and migration progress; /unipi:memory migrate converts v2 data after a confirmed plan; recall on|off / write on|off are session-only.

Naming Convention

Format: <most_important>_<less_important>_<lesser> — underscores, specific, <60 chars (auth_jwt_prefer_refresh_tokens, not auth_tokens).

Guardrails

• Read max 10 results per search.
• Update existing memories — the store tool detects exact titles and lists

similar ones.

• v2 leftovers are migrated by the user-initiated /unipi:memory migrate

(backups first); check progress via /unipi:memory status.