uteke-pi

Uteke persistent memory extensions for the pi coding agent.

uteke-remote makes uteke your agent's long-term memory: it auto-recalls relevant memories before each turn, can auto-save key facts after each turn, and registers uteke_remember / uteke_recall tools the LLM can call on demand — all over plain HTTP against any uteke server (local uteke-serve or remote). No CLI spawning, no local store required.

Features

• Auto-recall — before every agent turn, semantic memories relevant to the user's prompt are injected as context (project-tag scoped, with namespace-wide fallback).
• Auto-save — after a turn, a condensed summary of the key takeaway is stored when it contains decision/fix/config/architecture signals; trivial responses are filtered. On by default (UTEKE_AUTOSAVE=false to disable).
• Direct tools — uteke_remember / uteke_recall for explicit LLM-driven memory operations (can be disabled to coexist with the uteke MCP server).
• Commands — /uteke, /uteke-recall, /uteke-save, /uteke-on, /uteke-off, /uteke-autosave.
• Project isolation — every save is tagged project:<name> and recalls are scoped to the current project first, preventing cross-project leakage in a shared namespace.
• Resilient — server unreachable? Everything degrades silently; agent flow is never blocked.

Install

pi install git:github.com/codecoradev/uteke-pi

Or try it once without installing:

pi -e git:github.com/codecoradev/uteke-pi

Configuration

All config is via environment variables (no credentials are stored in this repo):

Env	Default	Description
UTEKE_BASE_URL	http://localhost:8767	uteke server base URL
UTEKE_TOKEN	(empty)	Bearer token (required when the server enforces auth)
UTEKE_NAMESPACE	default	Memory namespace (multi-agent isolation)
UTEKE_RECALL_LIMIT	5	Max memories injected per turn
UTEKE_MIN_SCORE	0.80	Similarity floor for injected memories
UTEKE_TIMEOUT_MS	8000	Per-request HTTP timeout
UTEKE_TOOLS	auto	auto (default) detects a uteke entry in pi's mcp.json (project .pi/mcp.json or global ~/.pi/agent/mcp.json) and skips direct tools when found; direct always registers tools; off always skips them
UTEKE_AUTOSAVE	true	default ON: after each turn, a condensed summary is stored when it contains decision/fix/config/architecture signals (trivial responses are filtered). Set false/0 to start disabled, or toggle at runtime with /uteke-autosave
UTEKE_PROJECT / UTEKE_PROJECT_TAG	(auto)	Override project tag detection

Example shell profile:

export UTEKE_BASE_URL="https://uteke.example.com"
export UTEKE_TOKEN="your-token-here"
export UTEKE_NAMESPACE="my-agent"

Running a server

• Local: uteke-serve from the uteke releases listens on 127.0.0.1:8767.
• Remote: any uteke server deployment works — the extension is just an HTTP client.

Using with the uteke MCP server

The MCP server exposes uteke_remember / uteke_recall / and more as MCP tools. If you use both, this is detected automatically: when a uteke server is configured in pi's mcp.json (project .pi/mcp.json or global ~/.pi/agent/mcp.json), the default UTEKE_TOOLS=auto skips registering the direct tools so names don't collide. Set UTEKE_TOOLS=direct to force both surfaces. Either way, the extension's auto-recall/auto-save/commands keep working, and both surfaces read/write the same data.

Commands

Command	Action
/uteke	Status: server health, totals, toggles
/uteke-recall <query>	Manual semantic recall
/uteke-save <text> [--tags a,b]	Store a memory
/uteke-on / /uteke-off	Toggle auto-recall injection
/uteke-autosave	Toggle auto-save

Server API used

Endpoint	Purpose
POST /remember	Store {content, tags[], namespace, type?} → {id}
POST /recall	Semantic search {query, namespace, limit, tags?} → [{memory, score}]
GET /health	Liveness + memory count
GET /stats	Totals for the status command

License

Apache-2.0 — same as uteke itself.