pi-memsearch

Long-term memory for pi: semantic recall over a memsearch store shared with Claude Code, Codex, OpenClaw and OpenCode

Packages

Package details

extensionskillprompt

Install pi-memsearch from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:pi-memsearch
Package
pi-memsearch
Version
1.4.3
Published
Sep 16, 2026
Downloads
1,392/mo · 138/wk
Author
sripwoud
License
MIT
Types
extension, skill, prompt
Size
120.8 KB
Dependencies
0 dependencies · 4 peers
Pi manifest JSON
{
  "skills": [
    "./skills"
  ],
  "prompts": [
    "./prompts"
  ],
  "extensions": [
    "./extensions/memsearch.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-memsearch

npm ci

pi-memsearch.espadat.com

pi ships no memory by design: "primitives, not features". This package gives it long-term memory through memsearch, the same per-project memory store Claude Code, Codex, OpenClaw and OpenCode already write to.

  • Recall in the phrasing you use weeks later: 32/35 strong hits vs 26/35 for pi-memory's qmd backend, over 35 queries against an identical 223-file corpus (benchmark).
  • Cross-agent: pi recalls what Claude Code learned yesterday in the same repo, and vice versa.
  • Plain markdown under .memsearch/, yours to commit or gitignore.

The memory writes itself, then answers weeks later:

# .memsearch/memory/2026-08-13.md  ← written by the session, unprompted
### 22:41
- the user and the agent moved the hot cache to Redis with 5 minute TTLs

# a new session, three weeks on
you ▸ /recall how did we fix the flaky redis test?
pi  ▸ memory_search → 5 chunks; top: 2026-08-13 "moved the hot cache to Redis with 5 minute TTLs" (0.81)
      memory_expand → the full "### 22:41" section, with its session anchor
      → answered at layer 2; the origin transcript was never opened

Install

Current release: 1.4.3.

Needs uv, the only external dependency, plus pi >= 0.84.1 and Node >= 22.19.

pi install npm:pi-memsearch

The first run downloads the onnx embedding model once, about 560 MB. pi announces it as a notice, so the pause is not mistaken for a hang. No API key is involved.

Project-local, pinned and unreleased installs, and what survives an uninstall: the install guide.

Documentation

Everything is at pi-memsearch.espadat.com:

  • Install: prerequisites, the four install forms, first run, uninstall
  • Configuration: every PI_MEMSEARCH_* variable, with its default and effect
  • Tools: the seven tools, and the three-rung recall ladder
  • Memory store: the daily markdown files, and which store a session writes to
  • Runtime: hook-by-hook behavior, every tunable, every degradation path
  • Troubleshooting: start with memory_status
  • Limits: what is deliberately unsupported, and how it compares to pi-memory
  • Development: the mise tasks, and the no-build-step jiti loading

In the repo: CONTEXT.md for the vocabulary, and meta/adr/ for the decisions and their rejected alternatives. Mesh parity (ADR 0001) constrains the rest.

Development

mise run setup, then mise run check / test / dev. Task reference, test layout and the release process: the development page and CONTRIBUTING.md.