omp packages

pi-memory-wiki-all-you-need0.26.1

Central memory + LLM wiki for pi: always-injected memory, per-turn LLM curation, wiki written from memory.

kht6163 Published 5 days ago MIT ~2.1K downloads/mo

Untested on omp

This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.

Did it work for you?

Result

One report per person. A new report replaces your old one.

Reports 0

No reports yet.

Readme

pi-memory-wiki-all-you-need

The pi extension for memory-wiki-all-you-need — a central memory + LLM wiki server for the pi coding agent. (한국어)

  • Injects memory into every request — global, user and current-project memories go into the system prompt before each turn, and related memories are recalled per prompt.
  • Curates after each turn — the finished turn is sent to the server, whose LLM adds, updates or removes memories.
  • Gives the agent tools — memory_search, session_search, memory_add, memory_replace, memory_remove, memory_graph, wiki_search, wiki_read, wiki_write, skill_manage, memory_review (read the memory review's pending proposals; settle a conflict after checking the real code), plus /memory, /memory-config, /memory-server, /memory-pin, /memory-flush, /wiki-compose, /skills-sync.
  • Brings the server's skills — on every pi start and with /skills-sync, the global and current-project skills written on the server are downloaded (one way; local edits are overwritten) to ~/.pi/agent/extensions/memory-wiki-all-you-need/skills/ and offered to pi as skills. The agent saves procedures itself with skill_manage (create/update on the server, never delete), and you are told on the next prompt when the server's skills changed.

Projects are identified by the git origin URL. The extension is fail-soft: if the server is down or slow, pi keeps going after a short timeout and reuses the last memory block it received.

You need a running server. See the main README for how to set one up.

Install

pi install npm:pi-memory-wiki-all-you-need

Then point it at your server once, inside pi:

/memory-server http://<server>:8765

This saves the URL to ~/.pi/agent/extensions/memory-wiki-all-you-need.json (under PI_CODING_AGENT_DIR if set), where other pi extensions keep their settings too. You can also write the file yourself:

{ "serverUrl": "http://<server>:8765" }

Run /memory-server without an argument to see the current URL and where it comes from.

If you already installed the extension with the server's /install.sh, delete the ~/.pi/agent/extensions/memory-wiki-all-you-need/ folder first (keep the .json settings file). With both installed, the extension runs twice.

Update with pi update.

Settings

All keys of memory-wiki-all-you-need.json are optional. Change them inside pi with /memory-config:

/memory-config                    # list the settings and pick one to change
/memory-config timeoutMs 3000     # set a value (applies right away)
/memory-config unset project      # remove a key (back to the default)

An environment variable with the same meaning overrides the file.

Key Env override Default Description
serverUrl MEMORY_SERVER_URL – (set with /memory-server) Server URL
settleDelayMs MEMORY_SETTLE_DELAY_MS 8000 Wait after a turn settles before sending it to the server
timeoutMs MEMORY_TIMEOUT_MS 1500 Timeout for fetching memory to inject
project MEMORY_PROJECT (auto) Override the project key
skillNudge MEMORY_SKILL_NUDGE 8 After a run with this many tool calls (2+ tools) and no skill_manage call, hint the agent to save a skill; 0 = off
showActivity MEMORY_SHOW_ACTIVITY true Show recalled memories (memory_recall) and curation results (memory_curate: added / updated / deleted) as tool-like cards in the conversation, one line per memory; ctrl+o expands them (full title and the start of each memory). Display only: what the model receives does not change
disabled MEMORY_DISABLED=1 false Disable the extension (/memory-config disabled false turns it back on)

License

MIT