@memongo/pi-extension

Pi coding-agent extension — exposes Memongo's durable, cross-project, MongoDB-backed semantic memory as additive tools alongside pi-hermes-memory.

Packages

Package details

extension

Install @memongo/pi-extension from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:@memongo/pi-extension
Package
@memongo/pi-extension
Version
2.2.2
Published
Oct 12, 2026
Downloads
230/mo · 19/wk
Author
romiluz13
License
Apache-2.0
Types
extension
Size
57.5 KB
Dependencies
2 dependencies · 3 peers
Pi manifest JSON
{
  "extensions": [
    "./extensions"
  ]
}

Security note

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

README

@memongo/pi-extension

Memongo durable memory for the Pi coding agent.

Additive — sits alongside pi-hermes-memory (local SQLite FTS5) and exposes Memongo's hybrid vector + full-text + graph retrieval as new tools. Does not replace any existing Pi memory tool.

Tools

Tool Purpose
memongo_search Semantic / cross-project / cross-session search over durable memory (hybrid vector + text + graph)
memongo_save Persist a durable structured fact / decision / preference / instruction to Memongo
memongo_status Probe Memongo API health + vector search availability
/memongo Slash command — quick status check

Quick start

1. Run Memongo locally

cd /path/to/memongo

# Optional automated semantic search (Preview; Atlas Model API key):
export VOYAGE_API_KEY="al-..."

# Start MongoDB (Atlas Local Preview — includes mongod + mongot + Atlas Search + auto-embed)
docker compose -f docker/docker-compose.yml up -d

# Start the HTTP API
cd apps/api && bun run dev

VOYAGE_API_KEY must be a MongoDB Atlas Model API key with the al-... prefix, not a direct Voyage pa-... key. MongoDB Automated Embedding is an upstream Preview feature that MongoDB says not to use in production.

2. Wire into Pi

Add to ~/.pi/agent/settings.json:

{
  "packages": [
    "/path/to/memongo/packages/pi-extension"
  ]
}

Set env vars (in shell or .envrc):

export MEMONGO_API_URL="http://127.0.0.1:3847"
export MEMONGO_API_KEY="local-dev-secret"
# optional: MEMONGO_AGENT_ID defaults to "pi" in this extension

3. Reload and verify

/reload
/memongo

You should see Memongo @ http://127.0.0.1:3847 with vector status. Then try:

  • memongo_status — confirms API + vector search are live
  • memongo_search with a query — semantic recall across projects
  • memongo_save — persist a durable fact

Architecture

Pi Agent
  ├── pi-hermes-memory (local FTS5)       ← stays
  ├── pi-observational-memory             ← stays
  └── @memongo/pi-extension (this)        ← additive
        └── @memongo/client → HTTP API (port 3847) → MongoDB (Docker)

Config

Env var Default Purpose
MEMONGO_API_URL http://127.0.0.1:3847 Memongo HTTP API base URL
MEMONGO_API_KEY (none) Bearer token (optional for local dev)
MEMONGO_AGENT_ID pi Agent identity for scoping
MEMONGO_PI_AUTO_CAPTURE 0 (off) Opt in to automatic capture of your turns (see below)
MEMONGO_PI_SESSION_INJECTION 1 (on) Inject recalled memory at session start
MEMONGO_PI_MEMORY_SCOPE agent Memory scope for capture and recall
VOYAGE_API_KEY (none) Required for vector/semantic search (Atlas Model API key)

Each surface has its own agentId default — pi here, main for the API bridge and web console — so Pi's memory stays a separate tenant from other surfaces by default. Set MEMONGO_AGENT_ID to share identity across surfaces; the web console's agent field views any agent regardless.

Data boundary (auto-capture)

Auto-capture (turn text sent to Memongo without an explicit tool call) is off by default. At registration the extension prints one notice stating this boundary. Opting in with MEMONGO_PI_AUTO_CAPTURE=1 sends, per turn:

  • the raw text of your user prompts and the agent's assistant replies — no redaction, no filtering
  • the session id, agent id, and resolved memory scope

Tool calls, tool results, images, and thinking blocks are not captured. Captured text is stored as canonical events and embedded for semantic recall; secrets redaction applies to logs and diagnostics, not stored content — treat anything captured as retained data subject to your retention/erasure settings. Explicit saves via memongo_save are unaffected by this switch.

Graceful degradation

If the Memongo API is down, all three tools return clean error messages instead of crashing. The agent is instructed to fall back to memory_search (local FTS5). Nothing breaks.

License

Apache-2.0