@industrialalgebra/ijima-pi

Ijima memory tools for pi — the Anima ecosystem's centralized agent memory service.

Registers nine tools backed by a wasm-mapped client core:

• memory_search, memory_save, memory_delete, memory_check_duplicate
• knowledge_add, knowledge_query, knowledge_status, knowledge_invalidate, knowledge_timeline

Per-project configuration (.pi/ijima.json)

A repo can declare which memory wall it belongs to — the declarative form of org scoping. Put a config at the project root:

{ "namespace": "ns_orthant_shared" }

Optional keys url and token_file pin the daemon and credentials the same way. Resolution walks up from the session root (nearest file wins), and precedence is always environment variable → project file → default, so an exported IJIMA_NAMESPACE still overrides the declaration. Invalid JSON is ignored silently — a broken config never breaks a session.

Setup

In ~/.pi/agent/settings.json:

{ "packages": ["npm:@industrialalgebra/ijima-pi"] }

Point it at your daemon with one multi-capability grant:

IJIMA_URL=http://your-ijima-host:7373
IJIMA_TOKEN=<grant from: ijima token issue --principal <name> \
  --capabilities memory:read,memory:write,knowledge:read,knowledge:write>

Both may also live in your shell environment. The grant needs one capability per tool family; a missing capability surfaces in the tool error with the exact ijima token issue invocation to fix it.

Building from source

npm run build   # wasm-pack (the ijima-pi crate) + tsc shim

Auto-capture and wake-up (the loop-closers)

The extension closes the memory loop without agent diligence:

• Auto-capture — after each assistant turn, the exchange is stored at the AutoCapture trust tier (importance 0.5, length-gated, truncated at 2000 chars, silent failure — capture never interrupts a session).
• Wake-up injection — every system prompt gains an ## Agent Memory (ACTIVE) block: a tool reminder plus your wake-up essentials and doctrine (GET /wakeup), refreshed per session.
• Token fallback — IJIMA_TOKEN env, then ~/.config/ijima/token (or $IJIMA_TOKEN_FILE), so shells that didn't source the env still work. IJIMA_URL falls back to ~/.config/ijima/url the same way.
• Home namespace — set IJIMA_NAMESPACE=ns_<org>_shared and captures, saves, and wake-up operate in that wall instead of the personal namespace: shared knowledge stays shared (and other orgs stay walled). Requires daemon ≥ 0.2.5 for namespaced wake-up; captures work on 0.2.2+.
• Content-derived ids — memories are id'd by content hash (mem_<hash16>): same content, same id, meaningful in transcripts.

memory_save remains the deliberate path: explicit saves land at the Explicit tier with higher importance — auto-capture is the floor, not the ceiling. An E2E harness (e2e.mjs, run against a live daemon) proves the full cycle: capture → wake-up → injection.

The bundled skill

The package ships a ijima skill (auto-registered by pi alongside the tools): the namespace mental model, why "empty" results are usually scoping rather than missing data, a diagnostics ladder, and memory-saving conventions. If an agent reports the brain looks empty, point it at that skill before it speculates about server state.

Production track record (stability report)

First production deployment: August 2026. Reviewed at six weeks of soak (2026-09-21). One central daemon, a seven-host agent fleet, ~38,700 memories across namespace walls at review time.

Signal	State at review
Uptime	All restarts operator-driven (config switches); zero crashes, zero auto-restarts
Error log	43 error lines total across the whole soak — all from one known incident (a database lock race against a backup drill, cause fixed and doctrine-encoded); error-free since
Capture volume	~90–180 new memories/day; growth is entirely agent captures + curated corpus
Release cadence	0.1.0 → 0.3.0 in two months; npm publishes via OIDC trusted publishing in CI
Backup discipline	Hourly mirror to a second filesystem, weekly restore drills that boot the mirror, daily census snapshots

The loop features (memory_search priming at wake-up, auto-capture at low trust, memory_save for deliberate persistence) have been running unattended across the fleet for the entire soak. Three documented organic-learning events from that period — an agent that learned a publishing calendar it was never told about, a research radar that identified a transferable architecture seam, and a cross-repo sprint planned from remembered dives on a different repository — are written up on the company blog.

The full soak log, runbooks, and restore-drill records live in the repository.

License

Apache-2.0 — same as the Ijima workspace.