pi

nt ↔ Pi — a memory, rules & knowledge-base system

This bundle makes nt the memory, rules, and knowledge-base backend for Pi (the minimal terminal coding agent), wired the way Pi is extended: an in-process extension, three prompt templates (/recall in, /learn out, /distill to consolidate), a skill, and a thin AGENTS.md. The agent's memory then survives across sessions, lives in plain files you can grep/git diff/open in Obsidian, and costs the right number of tokens for each kind of content.

nt pi install   # complete setup from any installed nt binary (--print to preview)

(From a repo checkout, ./install.sh does the same.)

Pi has no MCP — so the extension is the bridge

The one structural difference from the OpenCode integration: Pi has no built-in MCP — its docs say "No MCP. …build an extension that adds MCP support." nt already ships an MCP server (nt mcp, newline-delimited JSON-RPC over stdio), so the nt-memory extension spawns it, lists its tools, and registers each as a native Pi tool (pi.registerTool). Names + schemas come straight from nt mcp, so nt_index, nt_recall, nt_note, … appear exactly as under any MCP client and stay in sync with the binary. NT_BRIDGE=0 runs injection-only (the agent then drives nt via the CLI over bash).

Three layers, matched to three Pi surfaces

The core problem is a token budget one: Pi's rules layer (AGENTS.md + SYSTEM.md) is static text billed on every request, so the question per kind of memory is "should it be in context all the time?" That splits three ways:

Keep the rules + core-memory core small; keep the bulk KB behind the tools. Promoting a reference note to a rule is a retag (nt_tag … +rule), never a copy.

Recall loop. Record a mistake as a lesson (nt_note tagged lesson, trigger in the description). At each task start the agent nt_recalls a plain-words description — paraphrase-aware, lessons ranked first. And a failed bash command auto-triggers a lessons-only recall whose hits are appended onto the result, so the mistake summons its own antidote next turn. Lessons cost tokens only when recalled.

What's in the bundle

• extensions/nt-memory.ts — the whole system, defensively wrapped so a broken nt never breaks a session:
  • Tool bridge — spawns nt mcp, handshakes, registerTools each nt tool (read: nt_index/nt_search/nt_recall/nt_get/nt_status/nt_links; write: nt_add/nt_note/nt_note_edit/nt_update/nt_tag/nt_mv/ nt_archive/nt_relink/nt_rm); torn down on session_shutdown. Since registerTool only runs once at load but the subprocess can die mid session (a crash, or session_shutdown firing on something short of a real end — e.g. /new or /fork), the bridge self-heals: the next bridged call lazily respawns it, with concurrent calls sharing one in-flight respawn and a cooldown bounding retries against a persistently broken nt.
  • Rules + core-memory injection every run (before_agent_start), recompiled live from nt export, capped at NT_INJECT_MAX and truncated on note boundaries (never mid-rule). Re-running each turn means rules also survive compaction. NT_INJECT=off disables.
  • Error-triggered recall (tool_result, NT_ERROR_RECALL=0 to disable).
  • Idle nudge (agent_end, NT_IDLE_NUDGE=0 to disable) — one toast per session suggesting /learn, after NT_IDLE_NUDGE_THRESHOLD (default 3) tool-using turns with nothing saved (not the first — a session that writes a few turns in shouldn't get nagged).
  • If the bridge can't start, the injected rules still apply and the agent falls back to the nt CLI.
• skills/nt/SKILL.md — the recall-first / capture-the-why workflow (/skill:nt).
• prompts/{learn,recall,distill}.md — human-gated session harvest, on-demand briefing, and store consolidation (Pi's bash-style $@ args). /distill runs two passes: pass 1 lists every near-dup pair via nt_distill (read-only) and walks each to approval before merging or tagging a deliberate fork distinct; pass 2 reviews the always-injected rules/+memory/ block for subsumption, contradictions, dead triggers, and rules that aren't actually always-relevant (demoted to lessons, not deleted). One approval list, nothing written without approval. /distill rules runs pass 2 alone.
• AGENTS.md — a thin nudge (Pi concatenates it from ~/.pi/agent/, parent dirs, and cwd). README.md, install.sh, embed.go.

Install & verify

nt pi install          # from any installed binary
nt pi install --print  # preview, change nothing

Or from a checkout: cd integrations/pi && ./install.sh (or NT_BIN=/abs/nt ./install.sh). Both are idempotent; re-running after an nt upgrade refreshes the files. Then restart Pi (or /reload) and verify with nt export --tag rule --title Rules — exactly what gets injected. In a session the agent should call nt_status/nt_search and act on a <nt-memory> block.

nt note "Always prefer table-driven tests" --kind rule --description "…"       # rule
nt note "User deploys via 'make ship'" --kind memory --description "…"         # core memory
nt note "Auth uses 24h JWTs, 7d refresh" --kind ref --tag auth --description "…"  # KB (on-demand)

Bracket a session with /recall <topic> at the start and /learn at the end.

Config & environment

Config dir ~/.pi/agent/ (override with PI_CODING_AGENT_DIR); files land in extensions/nt-memory.ts, skills/nt/SKILL.md, prompts/{learn,recall,distill}.md, AGENTS.md. nt must be on Pi's PATH, or set NT_BIN.

Requirements

nt on PATH (or NT_BIN). Pi with the extension API (registerTool, pi.on). The bridge spawns nt mcp; if it can't start, injection and the CLI fallback still work.